Skip to content

V
VMAlgorithms_FullPaper_AAAI25
Commit 60bf36cc by GongYu
Options

新版

parent 9c95a8b3
Showing with 1862 additions and 1062 deletions
Apendix/anonymous-submission-latex-2024.aux
\relax \relax
\bibstyle{aaai24} \bibstyle{aaai24}
\citation{borkar1997stochastic}
\citation{hirsch1989convergent}
\citation{borkar2000ode}
\citation{borkar2000ode}
\citation{borkar2000ode}
\newlabel{proofth1}{{A.1}{1}}
\newlabel{th1proof}{{A.1}{1}}
\newlabel{thetaFast}{{A-1}{1}}
\newlabel{omegaFast}{{A-2}{1}}
\newlabel{omegaFastFinal}{{A-3}{1}}
\newlabel{omegaInfty}{{A-4}{1}}
\citation{sutton2009fast} \citation{sutton2009fast}
\newlabel{odetheta}{{A-5}{2}}
\newlabel{covariance}{{A-6}{2}}
\newlabel{odethetafinal}{{A-7}{2}}
\newlabel{proofth2}{{A.2}{2}}
\citation{hirsch1989convergent} \citation{hirsch1989convergent}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{hirsch1989convergent} \citation{hirsch1989convergent}
\newlabel{proofth2}{{A.1}{1}}
\newlabel{thetavmtdcFastest}{{A-1}{1}}
\newlabel{uvmtdcFastest}{{A-2}{1}}
\newlabel{omegavmtdcFastest}{{A-3}{1}}
\newlabel{omegavmtdcFastestFinal}{{A-4}{1}}
\newlabel{omegavmtdcInfty}{{A-5}{1}}
\newlabel{thetavmtdcFaster}{{A-6}{1}}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\newlabel{thetavmtdcFastest}{{A-8}{3}}
\newlabel{uvmtdcFastest}{{A-9}{3}}
\newlabel{omegavmtdcFastest}{{A-10}{3}}
\newlabel{omegavmtdcFastestFinal}{{A-11}{3}}
\newlabel{omegavmtdcInfty}{{A-12}{3}}
\newlabel{thetavmtdcFaster}{{A-13}{3}}
\newlabel{uvmtdcFaster}{{A-14}{3}}
\newlabel{uvmtdcFasterFinal}{{A-15}{3}}
\newlabel{uvmtdcInfty}{{A-16}{3}}
\citation{borkar1997stochastic} \citation{borkar1997stochastic}
\newlabel{uvmtdcFaster}{{A-7}{2}}
\newlabel{uvmtdcFasterFinal}{{A-8}{2}}
\newlabel{uvmtdcInfty}{{A-9}{2}}
\newlabel{thetavmtdcSlowerFinal}{{A-11}{2}}
\newlabel{odethetavmtdcfinal}{{A-12}{2}}
\citation{hirsch1989convergent} \citation{hirsch1989convergent}
\newlabel{thetavmtdcSlowerFinal}{{A-18}{4}}
\newlabel{odethetavmtdcfinal}{{A-19}{4}}
\newlabel{proofVMETD}{{A.3}{4}}
\newlabel{th1proof}{{A.3}{4}}
\newlabel{thetaFast}{{A-20}{4}}
\newlabel{omegaFast}{{A-21}{4}}
\newlabel{omegaFastFinal}{{A-22}{4}}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\citation{borkar2000ode} \citation{borkar2000ode}
\newlabel{proofVMETD}{{A.2}{3}} \newlabel{omegaInfty}{{A-23}{5}}
\newlabel{th1proof}{{A.2}{3}} \newlabel{odetheta}{{A-24}{5}}
\newlabel{thetaFast}{{A-13}{3}}
\newlabel{omegaFast}{{A-14}{3}}
\newlabel{omegaFastFinal}{{A-15}{3}}
\newlabel{omegaInfty}{{A-16}{3}}
\citation{sutton2016emphatic} \citation{sutton2016emphatic}
\newlabel{odetheta}{{A-17}{4}} \newlabel{rowsum}{{A-27}{6}}
\newlabel{rowsum}{{A-20}{4}} \newlabel{columnsum}{{A-28}{6}}
\citation{baird1995residual,sutton2009fast} \newlabel{odethetafinal}{{A-29}{6}}
\citation{baird1995residual,sutton2009fast,maei2011gradient} \newlabel{experimentaldetails}{{B}{6}}
\providecommand*\caption@xref[2]{\@setref\relax\@undefined{#1}}
\newlabel{bairdexample}{{1}{5}}
\newlabel{columnsum}{{A-21}{5}}
\newlabel{odethetafinal}{{A-22}{5}}
\newlabel{experimentaldetails}{{B}{5}}
\bibdata{aaai24} \bibdata{aaai24}
\bibcite{baird1995residual}{{1}{1995}{{Baird et~al.}}{{}}} \bibcite{borkar1997stochastic}{{1}{1997}{{Borkar}}{{}}}
\bibcite{borkar1997stochastic}{{2}{1997}{{Borkar}}{{}}} \bibcite{borkar2000ode}{{2}{2000}{{Borkar and Meyn}}{{}}}
\bibcite{borkar2000ode}{{3}{2000}{{Borkar and Meyn}}{{}}} \bibcite{hirsch1989convergent}{{3}{1989}{{Hirsch}}{{}}}
\bibcite{hirsch1989convergent}{{4}{1989}{{Hirsch}}{{}}} \bibcite{sutton2009fast}{{4}{2009}{{Sutton et~al.}}{{Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}}}
\bibcite{maei2011gradient}{{5}{2011}{{Maei}}{{}}} \bibcite{sutton2016emphatic}{{5}{2016}{{Sutton, Mahmood, and White}}{{}}}
\bibcite{sutton2009fast}{{6}{2009}{{Sutton et~al.}}{{Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}}} \providecommand*\caption@xref[2]{\@setref\relax\@undefined{#1}}
\bibcite{sutton2016emphatic}{{7}{2016}{{Sutton, Mahmood, and White}}{{}}} \newlabel{lrofways}{{1}{7}}
\newlabel{lrofways}{{1}{6}} \gdef \@abspage@last{7}
\gdef \@abspage@last{6}
Apendix/anonymous-submission-latex-2024.bbl
\begin{thebibliography}{7} \begin{thebibliography}{5}
\providecommand{\natexlab}[1]{#1} \providecommand{\natexlab}[1]{#1}
\bibitem[{Baird et~al.(1995)}]{baird1995residual}
Baird, L.; et~al. 1995.
\newblock Residual algorithms: Reinforcement learning with function approximation.
\newblock In \emph{Proc. 12th Int. Conf. Mach. Learn.}, 30--37.
\bibitem[{Borkar(1997)}]{borkar1997stochastic} \bibitem[{Borkar(1997)}]{borkar1997stochastic}
Borkar, V.~S. 1997. Borkar, V.~S. 1997.
\newblock Stochastic approximation with two time scales. \newblock Stochastic approximation with two time scales.
...@@ -21,11 +16,6 @@ Hirsch, M.~W. 1989. ...@@ -21,11 +16,6 @@ Hirsch, M.~W. 1989.
\newblock Convergent activation dynamics in continuous time networks. \newblock Convergent activation dynamics in continuous time networks.
\newblock \emph{Neural Netw.}, 2(5): 331--349. \newblock \emph{Neural Netw.}, 2(5): 331--349.
\bibitem[{Maei(2011)}]{maei2011gradient}
Maei, H.~R. 2011.
\newblock \emph{Gradient temporal-difference learning algorithms}.
\newblock Ph.D. thesis, University of Alberta.
\bibitem[{Sutton et~al.(2009)Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}]{sutton2009fast} \bibitem[{Sutton et~al.(2009)Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}]{sutton2009fast}
Sutton, R.; Maei, H.; Precup, D.; Bhatnagar, S.; Silver, D.; Szepesv{\'a}ri, C.; and Wiewiora, E. 2009. Sutton, R.; Maei, H.; Precup, D.; Bhatnagar, S.; Silver, D.; Szepesv{\'a}ri, C.; and Wiewiora, E. 2009.
\newblock Fast gradient-descent methods for temporal-difference learning with linear function approximation. \newblock Fast gradient-descent methods for temporal-difference learning with linear function approximation.
......
Apendix/anonymous-submission-latex-2024.blg
...@@ -3,44 +3,44 @@ Capacity: max_strings=200000, hash_size=200000, hash_prime=170003 ...@@ -3,44 +3,44 @@ Capacity: max_strings=200000, hash_size=200000, hash_prime=170003
The top-level auxiliary file: anonymous-submission-latex-2024.aux The top-level auxiliary file: anonymous-submission-latex-2024.aux
The style file: aaai24.bst The style file: aaai24.bst
Database file #1: aaai24.bib Database file #1: aaai24.bib
You've used 7 entries, You've used 5 entries,
2840 wiz_defined-function locations, 2840 wiz_defined-function locations,
630 strings with 5707 characters, 619 strings with 5446 characters,
and the built_in function-call counts, 4424 in all, are: and the built_in function-call counts, 3370 in all, are:
= -- 372 = -- 277
> -- 189 > -- 153
< -- 0 < -- 0
+ -- 74 + -- 60
- -- 64 - -- 52
* -- 295 * -- 242
:= -- 731 := -- 547
add.period$ -- 28 add.period$ -- 20
call.type$ -- 7 call.type$ -- 5
change.case$ -- 49 change.case$ -- 36
chr.to.int$ -- 8 chr.to.int$ -- 6
cite$ -- 7 cite$ -- 5
duplicate$ -- 302 duplicate$ -- 223
empty$ -- 320 empty$ -- 240
format.name$ -- 75 format.name$ -- 60
if$ -- 861 if$ -- 649
int.to.chr$ -- 1 int.to.chr$ -- 1
int.to.str$ -- 1 int.to.str$ -- 1
missing$ -- 63 missing$ -- 49
newline$ -- 39 newline$ -- 29
num.names$ -- 28 num.names$ -- 20
pop$ -- 125 pop$ -- 92
preamble$ -- 1 preamble$ -- 1
purify$ -- 45 purify$ -- 34
quote$ -- 0 quote$ -- 0
skip$ -- 134 skip$ -- 96
stack$ -- 0 stack$ -- 0
substring$ -- 246 substring$ -- 200
swap$ -- 160 swap$ -- 128
text.length$ -- 0 text.length$ -- 0
text.prefix$ -- 0 text.prefix$ -- 0
top$ -- 0 top$ -- 0
type$ -- 63 type$ -- 45
warning$ -- 0 warning$ -- 0
while$ -- 42 while$ -- 31
width$ -- 0 width$ -- 0
write$ -- 94 write$ -- 68
Apendix/anonymous-submission-latex-2024.log
This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.3.31) 12 AUG 2024 17:11 This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.3.31) 14 AUG 2024 06:25
entering extended mode entering extended mode
restricted \write18 enabled. restricted \write18 enabled.
file:line:error style messages enabled. file:line:error style messages enabled.
...@@ -582,7 +582,29 @@ File: ot1ptm.fd 2001/06/04 font definitions for OT1/ptm. ...@@ -582,7 +582,29 @@ File: ot1ptm.fd 2001/06/04 font definitions for OT1/ptm.
File: l3backend-pdftex.def 2023-01-16 L3 backend support: PDF output (pdfTeX) File: l3backend-pdftex.def 2023-01-16 L3 backend support: PDF output (pdfTeX)
\l__color_backend_stack_int=\count335 \l__color_backend_stack_int=\count335
\l__pdf_internal_box=\box82 \l__pdf_internal_box=\box82
) (./anonymous-submission-latex-2024.aux) ) (./anonymous-submission-latex-2024.aux
LaTeX Warning: Label `th1proof' multiply defined.
LaTeX Warning: Label `thetaFast' multiply defined.
LaTeX Warning: Label `omegaFast' multiply defined.
LaTeX Warning: Label `omegaFastFinal' multiply defined.
LaTeX Warning: Label `omegaInfty' multiply defined.
LaTeX Warning: Label `odetheta' multiply defined.
LaTeX Warning: Label `odethetafinal' multiply defined.
)
\openout1 = `anonymous-submission-latex-2024.aux'. \openout1 = `anonymous-submission-latex-2024.aux'.
LaTeX Font Info: Checking defaults for OML/cmm/m/it on input line 183. LaTeX Font Info: Checking defaults for OML/cmm/m/it on input line 183.
...@@ -627,41 +649,46 @@ Package caption Info: listings package is loaded. ...@@ -627,41 +649,46 @@ Package caption Info: listings package is loaded.
Package caption Info: End \AtBeginDocument code. Package caption Info: End \AtBeginDocument code.
Package newfloat Info: `float' package detected. Package newfloat Info: `float' package detected.
\c@lstlisting=\count342 \c@lstlisting=\count342
LaTeX Font Info: Trying to load font information for U+msa on input line 234. LaTeX Font Info: Trying to load font information for U+msa on input line 206.
(d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsa.fd (d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsa.fd
File: umsa.fd 2013/01/14 v3.01 AMS symbols A File: umsa.fd 2013/01/14 v3.01 AMS symbols A
) )
LaTeX Font Info: Trying to load font information for U+msb on input line 234. LaTeX Font Info: Trying to load font information for U+msb on input line 206.
(d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsb.fd (d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsb.fd
File: umsb.fd 2013/01/14 v3.01 AMS symbols B File: umsb.fd 2013/01/14 v3.01 AMS symbols B
) )
LaTeX Font Info: Trying to load font information for U+esvect on input line 234. LaTeX Font Info: Trying to load font information for U+esvect on input line 206.
(d:/software/texlive/2023/texmf-dist/tex/latex/esvect/uesvect.fd (d:/software/texlive/2023/texmf-dist/tex/latex/esvect/uesvect.fd
File: uesvect.fd File: uesvect.fd
) [1 )
LaTeX Warning: Reference `omega' on page 1 undefined on input line 265.
[1
{d:/software/texlive/2023/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{d:/software/texlive/2023/texmf-dist/fonts/enc/dvips/base/8r.enc}] [2] [3] [4] [5] [6] (./anonymous-submission-latex-2024.bbl) [7] (./anonymous-submission-latex-2024.aux)
LaTeX Warning: There were undefined references.
LaTeX Warning: There were multiply-defined labels.
{d:/software/texlive/2023/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{d:/software/texlive/2023/texmf-dist/fonts/enc/dvips/base/8r.enc}] [2] [3] [4] (./pic/BairdExample.tex) )
<pic/maze_13_13.pdf, id=34, 493.1646pt x 387.62602pt>
File: pic/maze_13_13.pdf Graphic file (type pdf)
<use pic/maze_13_13.pdf>
Package pdftex.def Info: pic/maze_13_13.pdf used on input line 902.
(pdftex.def) Requested size: 172.61018pt x 135.67113pt.
[5] (./anonymous-submission-latex-2024.bbl) [6 <./pic/maze_13_13.pdf>] (./anonymous-submission-latex-2024.aux) )
Here is how much of TeX's memory you used: Here is how much of TeX's memory you used:
22926 strings out of 476025 22606 strings out of 476025
482831 string characters out of 5789524 476412 string characters out of 5789524
1878382 words of memory out of 5000000 1879382 words of memory out of 5000000
43000 multiletter control sequences out of 15000+600000 42668 multiletter control sequences out of 15000+600000
531474 words of font info for 71 fonts, out of 8000000 for 9000 539762 words of font info for 95 fonts, out of 8000000 for 9000
1141 hyphenation exceptions out of 8191 1141 hyphenation exceptions out of 8191
84i,22n,89p,423b,789s stack positions out of 10000i,1000n,20000p,200000b,200000s 84i,22n,89p,423b,526s stack positions out of 10000i,1000n,20000p,200000b,200000s
<d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmib10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cmextra/cmmib7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmb8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmr8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmri8a.pfb> <d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmb8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmr8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmri8a.pfb>
Output written on anonymous-submission-latex-2024.pdf (6 pages, 200712 bytes). Output written on anonymous-submission-latex-2024.pdf (7 pages, 208835 bytes).
PDF statistics: PDF statistics:
110 PDF objects out of 1000 (max. 8388607) 117 PDF objects out of 1000 (max. 8388607)
68 compressed objects within 1 object stream 73 compressed objects within 1 object stream
0 named destinations out of 1000 (max. 500000) 0 named destinations out of 1000 (max. 500000)
18 words of extra memory for PDF output out of 10000 (max. 10000000) 13 words of extra memory for PDF output out of 10000 (max. 10000000)
Apendix/anonymous-submission-latex-2024.tex
...@@ -187,122 +187,287 @@ ...@@ -187,122 +187,287 @@
\onecolumn \onecolumn
\appendix \appendix
\section{Relevant proofs} \section{Relevant proofs}
% \subsection{VMTD} \subsection{Proof of Theorem 1}
% \begin{equation} \label{proofth1}
% \begin{array}{ccl} \begin{proof}
% \text{VBE}(\bm{\theta})&=&\mathbb{E}[(\mathbb{E}[\delta|s]-\kappa \mathbb{E}[\mathbb{E}[\delta|s]])^2]. \label{th1proof}
% \end{array} The proof is based on Borkar's Theorem for
% \end{equation} general stochastic approximation recursions with two time scales
\cite{borkar1997stochastic}.
% semi-gradient:
% \begin{equation}
% \begin{array}{ccl}
% 0&=&\mathbb{E}[\mathbb{E}[\delta|s]-\kappa \mathbb{E}[\mathbb{E}[\delta|s]](\bm{\phi} - \kappa\mathbb{E}[\bm{\phi}])]\\
% &=&\mathbb{E}[\delta \phi] - (2\kappa - \kappa^{2})\mathbb{E}[\delta]\mathbb{E}[\phi].
% \end{array}
% \end{equation}
% or
% \begin{equation}
% \begin{array}{ccl}
% 0&=&\mathbb{E}[\delta \phi] - \kappa\mathbb{E}[\delta]\mathbb{E}[\phi].
% \end{array}
% \end{equation}
% Therefore:
% \begin{equation}
% \begin{array}{ccl}
% \textbf{A}_{\text{VMTD}}&=&{\bm{\Phi}}^{\top} (\textbf{D}_{\mu}-(2\kappa - \kappa^{2})\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){\bm{\Phi}}.
% \end{array}
% \end{equation}
% or
% \begin{equation}
% \begin{array}{ccl}
% \textbf{A}_{\text{VMTD}}&=&{\bm{\Phi}}^{\top} (\textbf{D}_{\mu}-\kappa\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){\bm{\Phi}}.
% \end{array}
% \end{equation}
% The new TD error for the linear setting is
% \begin{equation*}
% \delta_{\text{new}}=r+\gamma
% \theta^{\top}\phi'-\theta^{\top}\phi-\mathbb{E}[\delta].
% \end{equation*}
A new one-step
linear TD solution is defined
as:
\begin{equation*}
0=\mathbb{E}[(\delta-\mathbb{E}[\delta]) \phi]=-A\theta+b.
\end{equation*}
Thus, the VMTD's solution is
$\theta_{\text{VMTD}}=A^{-1}b$.
First, note that recursion (5) can be rewritten as
\begin{equation*}
\theta_{k+1}\leftarrow \theta_k+\beta_k\xi(k),
\end{equation*}
where
\begin{equation*}
\xi(k)=\frac{\alpha_k}{\beta_k}(\delta_k-\omega_k)\phi_k
\end{equation*}
Due to the settings of step-size schedule $\alpha_k = o(\beta_k)$,
$\xi(k)\rightarrow 0$ almost surely as $k\rightarrow\infty$.
That is the increments in iteration (4) are uniformly larger than
those in (5), thus (4) is the faster recursion.
Along the faster time scale, iterations of (4) and (5)
are associated to ODEs system as follows:
\begin{equation}
\dot{\theta}(t) = 0,
\label{thetaFast}
\end{equation}
\begin{equation}
\dot{\omega}(t)=\mathbb{E}[\delta_t|\theta(t)]-\omega(t).
\label{omegaFast}
\end{equation}
Based on the ODE (\ref{thetaFast}), $\theta(t)\equiv \theta$ when
viewed from the faster timescale.
By the Hirsch lemma \cite{hirsch1989convergent}, it follows that
$||\theta_k-\theta||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
$\theta$ that depends on the initial condition $\theta_0$ of recursion
(5).
Thus, the ODE pair (\ref{thetaFast})-(\ref{omegaFast}) can be written as
\begin{equation}
\dot{\omega}(t)=\mathbb{E}[\delta_t|\theta]-\omega(t).
\label{omegaFastFinal}
\end{equation}
Consider the function $h(\omega)=\mathbb{E}[\delta|\theta]-\omega$,
i.e., the driving vector field of the ODE (\ref{omegaFastFinal}).
It is easy to find that the function $h$ is Lipschitz with coefficient
$-1$.
Let $h_{\infty}(\cdot)$ be the function defined by
$h_{\infty}(\omega)=\lim_{x\rightarrow \infty}\frac{h(x\omega)}{x}$.
Then $h_{\infty}(\omega)= -\omega$, is well-defined.
For (\ref{omegaFastFinal}), $\omega^*=\mathbb{E}[\delta|\theta]$
is the unique globally asymptotically stable equilibrium.
For the ODE
\begin{equation}
\dot{\omega}(t) = h_{\infty}(\omega(t))= -\omega(t),
\label{omegaInfty}
\end{equation}
apply $\vec{V}(\omega)=(-\omega)^{\top}(-\omega)/2$ as its
associated strict Liapunov function. Then,
the origin of (\ref{omegaInfty}) is a globally asymptotically stable
equilibrium.
\subsection{Proof of Theorem 1}
Consider now the recursion (\ref{omega}).
Let
$M_{k+1}=(\delta_k-\omega_k)
-\mathbb{E}[(\delta_k-\omega_k)|\mathcal{F}(k)]$,
where $\mathcal{F}(k)=\sigma(\omega_l,\theta_l,l\leq k;\phi_s,\phi_s',r_s,s<k)$,
$k\geq 1$ are the sigma fields
generated by $\omega_0,\theta_0,\omega_{l+1},\theta_{l+1},\phi_l,\phi_l'$,
$0\leq l<k$.
It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$.
Because $\phi_k$, $r_k$, and $\phi_k'$ have
uniformly bounded second moments, it can be seen that for some constant
$c_1>0$, $\forall k\geq0$,
\begin{equation*}
\mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq
c_1(1+||\omega_k||^2+||\theta_k||^2).
\end{equation*}
Now Assumptions (A1) and (A2) of \cite{borkar2000ode} are verified.
Furthermore, Assumptions (TS) of \cite{borkar2000ode} is satisfied by our
conditions on the step-size sequences $\alpha_k$, $\beta_k$. Thus,
by Theorem 2.2 of \cite{borkar2000ode} we obtain that
$||\omega_k-\omega^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$.
Consider now the slower time scale recursion (5).
Based on the above analysis, (5) can be rewritten as
\begin{equation*}
\theta_{k+1}\leftarrow
\theta_{k}+\alpha_k(\delta_k-\mathbb{E}[\delta_k|\theta_k])\phi_k.
\end{equation*}
Let $\mathcal{G}(k)=\sigma(\theta_l,l\leq k;\phi_s,\phi_s',r_s,s<k)$,
$k\geq 1$ be the sigma fields
generated by $\theta_0,\theta_{l+1},\phi_l,\phi_l'$,
$0\leq l<k$.
Let
$
Z_{k+1} = Y_{t}-\mathbb{E}[Y_{t}|\mathcal{G}(k)],
$
where
\begin{equation*}
Y_{k}=(\delta_k-\mathbb{E}[\delta_k|\theta_k])\phi_k.
\end{equation*}
Consequently,
\begin{equation*}
\begin{array}{ccl}
\mathbb{E}[Y_t|\mathcal{G}(k)]&=&\mathbb{E}[(\delta_k-\mathbb{E}[\delta_k|\theta_k])\phi_k|\mathcal{G}(k)]\\
&=&\mathbb{E}[\delta_k\phi_k|\theta_k]
-\mathbb{E}[\mathbb{E}[\delta_k|\theta_k]\phi_k]\\
&=&\mathbb{E}[\delta_k\phi_k|\theta_k]
-\mathbb{E}[\delta_k|\theta_k]\mathbb{E}[\phi_k]\\
&=&\mathrm{Cov}(\delta_k|\theta_k,\phi_k),
\end{array}
\end{equation*}
where $\mathrm{Cov}(\cdot,\cdot)$ is a covariance operator.
Thus,
\begin{equation*}
\begin{array}{ccl}
Z_{k+1}&=&(\delta_k-\mathbb{E}[\delta_k|\theta_k])\phi_k-\mathrm{Cov}(\delta_k|\theta_k,\phi_k).
\end{array}
\end{equation*}
It is easy to verify that $Z_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[Z_{k+1}|\mathcal{G}(k)]=0$, $\forall k\geq0$.
Also, because $\phi_k$, $r_k$, and $\phi_k'$ have
uniformly bounded second moments, it can be seen that for some constant
$c_2>0$, $\forall k\geq0$,
\begin{equation*}
\mathbb{E}[||Z_{k+1}||^2|\mathcal{G}(k)]\leq
c_2(1+||\theta_k||^2).
\end{equation*}
Consider now the following ODE associated with (5):
\begin{equation}
\begin{array}{ccl}
\dot{\theta}(t)&=&\mathrm{Cov}(\delta|\theta(t),\phi)\\
&=&\mathrm{Cov}(r+(\gamma\phi'-\phi)^{\top}\theta(t),\phi)\\
&=&\mathrm{Cov}(r,\phi)-\mathrm{Cov}(\theta(t)^{\top}(\phi-\gamma\phi'),\phi)\\
&=&\mathrm{Cov}(r,\phi)-\theta(t)^{\top}\mathrm{Cov}(\phi-\gamma\phi',\phi)\\
&=&\mathrm{Cov}(r,\phi)-\mathrm{Cov}(\phi-\gamma\phi',\phi)^{\top}\theta(t)\\
&=&\mathrm{Cov}(r,\phi)-\mathrm{Cov}(\phi,\phi-\gamma\phi')\theta(t)\\
&=&-A\theta(t)+b.
\end{array}
\label{odetheta}
\end{equation}
Let $\vec{h}(\theta(t))$ be the driving vector field of the ODE
(\ref{odetheta}).
\begin{equation*}
\vec{h}(\theta(t))=-A\theta(t)+b.
\end{equation*}
Consider the cross-covariance matrix,
\begin{equation}
\begin{array}{ccl}
A &=& \mathrm{Cov}(\phi,\phi-\gamma\phi')\\
&=&\frac{\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')-\mathrm{Cov}(\gamma\phi',\gamma\phi')}{2}\\
&=&\frac{\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')-\gamma^2\mathrm{Cov}(\phi',\phi')}{2}\\
&=&\frac{(1-\gamma^2)\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')}{2},\\
\end{array}
\label{covariance}
\end{equation}
where we eventually used $\mathrm{Cov}(\phi',\phi')=\mathrm{Cov}(\phi,\phi)$
\footnote{The covariance matrix $\mathrm{Cov}(\phi',\phi')$ is equal to
the covariance matrix $\mathrm{Cov}(\phi,\phi)$ if the initial state is re-reachable or
initialized randomly in a Markov chain for on-policy update.}.
Note that the covariance matrix $\mathrm{Cov}(\phi,\phi)$ and
$\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')$ are semi-positive
definite. Then, the matrix $A$ is semi-positive definite because $A$ is
linearly combined by two positive-weighted semi-positive definite matrice
(\ref{covariance}).
Furthermore, $A$ is nonsingular due to the assumption.
Hence, the cross-covariance matrix $A$ is positive definite.
Therefore,
$\theta^*=A^{-1}b$ can be seen to be the unique globally asymptotically
stable equilibrium for ODE (\ref{odetheta}).
Let $\vec{h}_{\infty}(\theta)=\lim_{r\rightarrow
\infty}\frac{\vec{h}(r\theta)}{r}$. Then
$\vec{h}_{\infty}(\theta)=-A\theta$ is well-defined.
Consider now the ODE
\begin{equation}
\dot{\theta}(t)=-A\theta(t).
\label{odethetafinal}
\end{equation}
The ODE (\ref{odethetafinal}) has the origin as its unique globally asymptotically stable equilibrium.
Thus, the assumption (A1) and (A2) are verified.
\end{proof}
\subsection{Proof of Theorem 2}
\label{proofth2} \label{proofth2}
\begin{proof} \begin{proof}
The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation. The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation.
For the VMTDC algorithm, a new one-step linear TD solution is defined as: For the VMTDC algorithm, a new one-step linear TD solution is defined as:
\begin{equation*} \begin{equation*}
0=\mathbb{E}[(\bm{\phi} - \gamma \bm{\phi}' - \mathbb{E}[\bm{\phi} - \gamma \bm{\phi}'])\bm{\phi}^\top]\mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta])\bm{\phi}]=\textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}\bm{\theta}+\bm{b}). 0=\mathbb{E}[({\phi} - \gamma {\phi}' - \mathbb{E}[{\phi} - \gamma {\phi}']){\phi}^\top]\mathbb{E}[{\phi} {\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta]){\phi}]=\textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}{\theta}+{b}).
\end{equation*} \end{equation*}
The matrix $\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}$ is positive definite. Thus, the VMTD's solution is The matrix $\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}$ is positive definite. Thus, the VMTD's solution is
$\bm{\theta}_{\text{VMTDC}}=\textbf{A}^{-1}\bm{b}$. ${\theta}_{\text{VMTDC}}=\textbf{A}^{-1}{b}$.
First, note that recursion (5) and (6) can be rewritten as, respectively, First, note that recursion (5) and (6) can be rewritten as, respectively,
\begin{equation*} \begin{equation*}
\bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\zeta_k \bm{x}(k), {\theta}_{k+1}\leftarrow {\theta}_k+\zeta_k {x}(k),
\end{equation*} \end{equation*}
\begin{equation*} \begin{equation*}
\bm{u}_{k+1}\leftarrow \bm{u}_k+\beta_k \bm{y}(k), {u}_{k+1}\leftarrow {u}_k+\beta_k {y}(k),
\end{equation*} \end{equation*}
where where
\begin{equation*} \begin{equation*}
\bm{x}(k)=\frac{\alpha_k}{\zeta_k}[(\delta_{k}- \omega_k) \bm{\phi}_k - \gamma\bm{\phi}'_{k}(\bm{\phi}^{\top}_k \bm{u}_k)], {x}(k)=\frac{\alpha_k}{\zeta_k}[(\delta_{k}- \omega_k) {\phi}_k - \gamma{\phi}'_{k}({\phi}^{\top}_k {u}_k)],
\end{equation*} \end{equation*}
\begin{equation*} \begin{equation*}
\bm{y}(k)=\frac{\zeta_k}{\beta_k}[\delta_{k}-\omega_k - \bm{\phi}^{\top}_k \bm{u}_k]\bm{\phi}_k. {y}(k)=\frac{\zeta_k}{\beta_k}[\delta_{k}-\omega_k - {\phi}^{\top}_k {u}_k]{\phi}_k.
\end{equation*} \end{equation*}
Recursion (5) can also be rewritten as Recursion (5) can also be rewritten as
\begin{equation*} \begin{equation*}
\bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\beta_k z(k), {\theta}_{k+1}\leftarrow {\theta}_k+\beta_k z(k),
\end{equation*} \end{equation*}
where where
\begin{equation*} \begin{equation*}
z(k)=\frac{\alpha_k}{\beta_k}[(\delta_{k}- \omega_k) \bm{\phi}_k - \gamma\bm{\phi}'_{k}(\bm{\phi}^{\top}_k \bm{u}_k)], z(k)=\frac{\alpha_k}{\beta_k}[(\delta_{k}- \omega_k) {\phi}_k - \gamma{\phi}'_{k}({\phi}^{\top}_k {u}_k)],
\end{equation*} \end{equation*}
Due to the settings of step-size schedule Due to the settings of step-size schedule
$\alpha_k = o(\zeta_k)$, $\zeta_k = o(\beta_k)$, $\bm{x}(k)\rightarrow 0$, $\bm{y}(k)\rightarrow 0$, $z(k)\rightarrow 0$ almost surely as $k\rightarrow 0$. $\alpha_k = o(\zeta_k)$, $\zeta_k = o(\beta_k)$, ${x}(k)\rightarrow 0$, ${y}(k)\rightarrow 0$, $z(k)\rightarrow 0$ almost surely as $k\rightarrow 0$.
That is that the increments in iteration (7) are uniformly larger than That is that the increments in iteration (7) are uniformly larger than
those in (6) and the increments in iteration (6) are uniformly larger than those in (6) and the increments in iteration (6) are uniformly larger than
those in (5), thus (7) is the fastest recursion, (6) is the second fast recursion and (5) is the slower recursion. those in (5), thus (7) is the fastest recursion, (6) is the second fast recursion and (5) is the slower recursion.
Along the fastest time scale, iterations of (5), (6) and (7) Along the fastest time scale, iterations of (5), (6) and (7)
are associated to ODEs system as follows: are associated to ODEs system as follows:
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t) = 0, \dot{{\theta}}(t) = 0,
\label{thetavmtdcFastest} \label{thetavmtdcFastest}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\dot{\bm{u}}(t) = 0, \dot{{u}}(t) = 0,
\label{uvmtdcFastest} \label{uvmtdcFastest}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\dot{\omega}(t)=\mathbb{E}[\delta_t|\bm{u}(t),\bm{\theta}(t)]-\omega(t). \dot{\omega}(t)=\mathbb{E}[\delta_t|{u}(t),{\theta}(t)]-\omega(t).
\label{omegavmtdcFastest} \label{omegavmtdcFastest}
\end{equation} \end{equation}
Based on the ODE (\ref{thetavmtdcFastest}) and (\ref{uvmtdcFastest}), both $\bm{\theta}(t)\equiv \bm{\theta}$ Based on the ODE (\ref{thetavmtdcFastest}) and (\ref{uvmtdcFastest}), both ${\theta}(t)\equiv {\theta}$
and $\bm{u}(t)\equiv \bm{u}$ when viewed from the fastest timescale. and ${u}(t)\equiv {u}$ when viewed from the fastest timescale.
By the Hirsch lemma \cite{hirsch1989convergent}, it follows that By the Hirsch lemma \cite{hirsch1989convergent}, it follows that
$||\bm{\theta}_k-\bm{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some $||{\theta}_k-{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
$\bm{\theta}$ that depends on the initial condition $\bm{\theta}_0$ of recursion ${\theta}$ that depends on the initial condition ${\theta}_0$ of recursion
(5) and $||\bm{u}_k-\bm{u}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some (5) and $||{u}_k-{u}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
$u$ that depends on the initial condition $u_0$ of recursion $u$ that depends on the initial condition $u_0$ of recursion
(6). Thus, the ODE pair (\ref{thetavmtdcFastest})-(ref{omegavmtdcFastest}) (6). Thus, the ODE pair (\ref{thetavmtdcFastest})-(ref{omegavmtdcFastest})
can be written as can be written as
\begin{equation} \begin{equation}
\dot{\omega}(t)=\mathbb{E}[\delta_t|\bm{u},\bm{\theta}]-\omega(t). \dot{\omega}(t)=\mathbb{E}[\delta_t|{u},{\theta}]-\omega(t).
\label{omegavmtdcFastestFinal} \label{omegavmtdcFastestFinal}
\end{equation} \end{equation}
Consider the function $h(\omega)=\mathbb{E}[\delta|\bm{\theta},\bm{u}]-\omega$, Consider the function $h(\omega)=\mathbb{E}[\delta|{\theta},{u}]-\omega$,
i.e., the driving vector field of the ODE (\ref{omegavmtdcFastestFinal}). i.e., the driving vector field of the ODE (\ref{omegavmtdcFastestFinal}).
It is easy to find that the function $h$ is Lipschitz with coefficient It is easy to find that the function $h$ is Lipschitz with coefficient
$-1$. $-1$.
Let $h_{\infty}(\cdot)$ be the function defined by Let $h_{\infty}(\cdot)$ be the function defined by
$h_{\infty}(\omega)=\lim_{r\rightarrow \infty}\frac{h(r\omega)}{r}$. $h_{\infty}(\omega)=\lim_{r\rightarrow \infty}\frac{h(r\omega)}{r}$.
Then $h_{\infty}(\omega)= -\omega$, is well-defined. Then $h_{\infty}(\omega)= -\omega$, is well-defined.
For (\ref{omegavmtdcFastestFinal}), $\omega^*=\mathbb{E}[\delta|\bm{\theta},\bm{u}]$ For (\ref{omegavmtdcFastestFinal}), $\omega^*=\mathbb{E}[\delta|{\theta},{u}]$
is the unique globally asymptotically stable equilibrium. is the unique globally asymptotically stable equilibrium.
For the ODE For the ODE
\begin{equation} \begin{equation}
...@@ -318,18 +483,18 @@ Consider now the recursion (7). ...@@ -318,18 +483,18 @@ Consider now the recursion (7).
Let Let
$M_{k+1}=(\delta_k-\omega_k) $M_{k+1}=(\delta_k-\omega_k)
-\mathbb{E}[(\delta_k-\omega_k)|\mathcal{F}(k)]$, -\mathbb{E}[(\delta_k-\omega_k)|\mathcal{F}(k)]$,
where $\mathcal{F}(k)=\sigma(\omega_l,\bm{u}_l,\bm{\theta}_l,l\leq k;\bm{\phi}_s,\bm{\phi}_s',r_s,s<k)$, where $\mathcal{F}(k)=\sigma(\omega_l,{u}_l,{\theta}_l,l\leq k;{\phi}_s,{\phi}_s',r_s,s<k)$,
$k\geq 1$ are the sigma fields $k\geq 1$ are the sigma fields
generated by $\omega_0,u_0,\bm{\theta}_0,\omega_{l+1},\bm{u}_{l+1},\bm{\theta}_{l+1},\bm{\phi}_l,\bm{\phi}_l'$, generated by $\omega_0,u_0,{\theta}_0,\omega_{l+1},{u}_{l+1},{\theta}_{l+1},{\phi}_l,{\phi}_l'$,
$0\leq l<k$. $0\leq l<k$.
It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$. satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$.
Because $\bm{\phi}_k$, $r_k$, and $\bm{\phi}_k'$ have Because ${\phi}_k$, $r_k$, and ${\phi}_k'$ have
uniformly bounded second moments, it can be seen that for some constant uniformly bounded second moments, it can be seen that for some constant
$c_1>0$, $\forall k\geq0$, $c_1>0$, $\forall k\geq0$,
\begin{equation*} \begin{equation*}
\mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq \mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq
c_1(1+||\omega_k||^2+||\bm{u}_k||^2+||\bm{\theta}_k||^2). c_1(1+||\omega_k||^2+||{u}_k||^2+||{\theta}_k||^2).
\end{equation*} \end{equation*}
...@@ -342,53 +507,53 @@ $||\omega_k-\omega^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$. ...@@ -342,53 +507,53 @@ $||\omega_k-\omega^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$.
Consider now the second time scale recursion (6). Consider now the second time scale recursion (6).
Based on the above analysis, (6) can be rewritten as Based on the above analysis, (6) can be rewritten as
% \begin{equation*} % \begin{equation*}
% \bm{u}_{k+1}\leftarrow u_{k}+\zeta_{k}[\delta_{k}-\mathbb{E}[\delta_k|\bm{u}_k,\bm{\theta}_k] - \bm{\phi}^{\top} (s_k) \bm{u}_k]\bm{\phi}(s_k). % {u}_{k+1}\leftarrow u_{k}+\zeta_{k}[\delta_{k}-\mathbb{E}[\delta_k|{u}_k,{\theta}_k] - {\phi}^{\top} (s_k) {u}_k]{\phi}(s_k).
% \end{equation*} % \end{equation*}
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t) = 0, \dot{{\theta}}(t) = 0,
\label{thetavmtdcFaster} \label{thetavmtdcFaster}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\dot{u}(t) = \mathbb{E}[(\delta_t-\mathbb{E}[\delta_t|\bm{u}(t),\bm{\theta}(t)])\bm{\phi}_t|\bm{\theta}(t)] - \textbf{C}\bm{u}(t). \dot{u}(t) = \mathbb{E}[(\delta_t-\mathbb{E}[\delta_t|{u}(t),{\theta}(t)]){\phi}_t|{\theta}(t)] - \textbf{C}{u}(t).
\label{uvmtdcFaster} \label{uvmtdcFaster}
\end{equation} \end{equation}
The ODE (\ref{thetavmtdcFaster}) suggests that $\bm{\theta}(t)\equiv \bm{\theta}$ (i.e., a time invariant parameter) The ODE (\ref{thetavmtdcFaster}) suggests that ${\theta}(t)\equiv {\theta}$ (i.e., a time invariant parameter)
when viewed from the second fast timescale. when viewed from the second fast timescale.
By the Hirsch lemma \cite{hirsch1989convergent}, it follows that By the Hirsch lemma \cite{hirsch1989convergent}, it follows that
$||\bm{\theta}_k-\bm{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some $||{\theta}_k-{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
$\bm{\theta}$ that depends on the initial condition $\bm{\theta}_0$ of recursion ${\theta}$ that depends on the initial condition ${\theta}_0$ of recursion
(5). (5).
Consider now the recursion (6). Consider now the recursion (6).
Let Let
$N_{k+1}=((\delta_k-\mathbb{E}[\delta_k]) - \bm{\phi}_k \bm{\phi}^{\top}_k \bm{u}_k) -\mathbb{E}[((\delta_k-\mathbb{E}[\delta_k]) - \bm{\phi}_k \bm{\phi}^{\top}_k \bm{u}_k)|\mathcal{I} (k)]$, $N_{k+1}=((\delta_k-\mathbb{E}[\delta_k]) - {\phi}_k {\phi}^{\top}_k {u}_k) -\mathbb{E}[((\delta_k-\mathbb{E}[\delta_k]) - {\phi}_k {\phi}^{\top}_k {u}_k)|\mathcal{I} (k)]$,
where $\mathcal{I}(k)=\sigma(\bm{u}_l,\bm{\theta}_l,l\leq k;\bm{\phi}_s,\bm{\phi}_s',r_s,s<k)$, where $\mathcal{I}(k)=\sigma({u}_l,{\theta}_l,l\leq k;{\phi}_s,{\phi}_s',r_s,s<k)$,
$k\geq 1$ are the sigma fields $k\geq 1$ are the sigma fields
generated by $\bm{u}_0,\bm{\theta}_0,\bm{u}_{l+1},\bm{\theta}_{l+1},\bm{\phi}_l,\bm{\phi}_l'$, generated by ${u}_0,{\theta}_0,{u}_{l+1},{\theta}_{l+1},{\phi}_l,{\phi}_l'$,
$0\leq l<k$. $0\leq l<k$.
It is easy to verify that $N_{k+1},k\geq0$ are integrable random variables that It is easy to verify that $N_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[N_{k+1}|\mathcal{I}(k)]=0$, $\forall k\geq0$. satisfy $\mathbb{E}[N_{k+1}|\mathcal{I}(k)]=0$, $\forall k\geq0$.
Because $\bm{\phi}_k$, $r_k$, and $\bm{\phi}_k'$ have Because ${\phi}_k$, $r_k$, and ${\phi}_k'$ have
uniformly bounded second moments, it can be seen that for some constant uniformly bounded second moments, it can be seen that for some constant
$c_2>0$, $\forall k\geq0$, $c_2>0$, $\forall k\geq0$,
\begin{equation*} \begin{equation*}
\mathbb{E}[||N_{k+1}||^2|\mathcal{I}(k)]\leq \mathbb{E}[||N_{k+1}||^2|\mathcal{I}(k)]\leq
c_2(1+||\bm{u}_k||^2+||\bm{\theta}_k||^2). c_2(1+||{u}_k||^2+||{\theta}_k||^2).
\end{equation*} \end{equation*}
Because $\bm{\theta}(t)\equiv \bm{\theta}$ from (\ref{thetavmtdcFaster}), the ODE pair (\ref{thetavmtdcFaster})-(\ref{uvmtdcFaster}) Because ${\theta}(t)\equiv {\theta}$ from (\ref{thetavmtdcFaster}), the ODE pair (\ref{thetavmtdcFaster})-(\ref{uvmtdcFaster})
can be written as can be written as
\begin{equation} \begin{equation}
\dot{\bm{u}}(t) = \mathbb{E}[(\delta_t-\mathbb{E}[\delta_t|\bm{\theta}])\bm{\phi}_t|\bm{\theta}] - \textbf{C}\bm{u}(t). \dot{{u}}(t) = \mathbb{E}[(\delta_t-\mathbb{E}[\delta_t|{\theta}]){\phi}_t|{\theta}] - \textbf{C}{u}(t).
\label{uvmtdcFasterFinal} \label{uvmtdcFasterFinal}
\end{equation} \end{equation}
Now consider the function $h(\bm{u})=\mathbb{E}[\delta_t-\mathbb{E}[\delta_t|\bm{\theta}]|\bm{\theta}] -\textbf{C}\bm{u}$, i.e., the Now consider the function $h({u})=\mathbb{E}[\delta_t-\mathbb{E}[\delta_t|{\theta}]|{\theta}] -\textbf{C}{u}$, i.e., the
driving vector field of the ODE (\ref{uvmtdcFasterFinal}). For (\ref{uvmtdcFasterFinal}), driving vector field of the ODE (\ref{uvmtdcFasterFinal}). For (\ref{uvmtdcFasterFinal}),
$\bm{u}^* = \textbf{C}^{-1}\mathbb{E}[(\delta-\mathbb{E}[\delta|\bm{\theta}])\bm{\phi}|\bm{\theta}]$ is the unique globally asymptotically ${u}^* = \textbf{C}^{-1}\mathbb{E}[(\delta-\mathbb{E}[\delta|{\theta}]){\phi}|{\theta}]$ is the unique globally asymptotically
stable equilibrium. Let $h_{\infty}(\bm{u})=-\textbf{C}\bm{u}$. stable equilibrium. Let $h_{\infty}({u})=-\textbf{C}{u}$.
For the ODE For the ODE
\begin{equation} \begin{equation}
\dot{\bm{u}}(t) = h_{\infty}(\bm{u}(t))= -\textbf{C}\bm{u}(t), \dot{{u}}(t) = h_{\infty}({u}(t))= -\textbf{C}{u}(t),
\label{uvmtdcInfty} \label{uvmtdcInfty}
\end{equation} \end{equation}
the origin of (\ref{uvmtdcInfty}) is a globally asymptotically stable the origin of (\ref{uvmtdcInfty}) is a globally asymptotically stable
...@@ -397,60 +562,60 @@ Now Assumptions (A1) and (A2) of \cite{borkar2000ode} are verified. ...@@ -397,60 +562,60 @@ Now Assumptions (A1) and (A2) of \cite{borkar2000ode} are verified.
Furthermore, Assumptions (TS) of \cite{borkar2000ode} is satisfied by our Furthermore, Assumptions (TS) of \cite{borkar2000ode} is satisfied by our
conditions on the step-size sequences $\alpha_k$,$\zeta_k$, $\beta_k$. Thus, conditions on the step-size sequences $\alpha_k$,$\zeta_k$, $\beta_k$. Thus,
by Theorem 2.2 of \cite{borkar2000ode} we obtain that by Theorem 2.2 of \cite{borkar2000ode} we obtain that
$||\bm{u}_k-\bm{u}^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$. $||{u}_k-{u}^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$.
Consider now the slower timescale recursion (5). In the light of the above, Consider now the slower timescale recursion (5). In the light of the above,
(5) can be rewritten as (5) can be rewritten as
\begin{equation} \begin{equation}
\bm{\theta}_{k+1} \leftarrow \bm{\theta}_{k} + \alpha_k (\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k\\ {\theta}_{k+1} \leftarrow {\theta}_{k} + \alpha_k (\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k\\
- \alpha_k \gamma\bm{\phi}'_{k}(\bm{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k])\bm{\phi}|\bm{\theta}_k]). - \alpha_k \gamma{\phi}'_{k}({\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]){\phi}|{\theta}_k]).
\end{equation} \end{equation}
Let $\mathcal{G}(k)=\sigma(\bm{\theta}_l,l\leq k;\bm{\phi}_s,\bm{\phi}_s',r_s,s<k)$, Let $\mathcal{G}(k)=\sigma({\theta}_l,l\leq k;{\phi}_s,{\phi}_s',r_s,s<k)$,
$k\geq 1$ be the sigma fields $k\geq 1$ be the sigma fields
generated by $\bm{\theta}_0,\bm{\theta}_{l+1},\bm{\phi}_l,\bm{\phi}_l'$, generated by ${\theta}_0,{\theta}_{l+1},{\phi}_l,{\phi}_l'$,
$0\leq l<k$. Let $0\leq l<k$. Let
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
Z_{k+1}&=&((\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k - \gamma \bm{\phi}'_{k}\bm{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k])\bm{\phi}|\bm{\theta}_k])\\ Z_{k+1}&=&((\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k - \gamma {\phi}'_{k}{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]){\phi}|{\theta}_k])\\
& &-\mathbb{E}[((\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k - \gamma \bm{\phi}'_{k}\bm{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k])\bm{\phi}|\bm{\theta}_k])|\mathcal{G}(k)]\\ & &-\mathbb{E}[((\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k - \gamma {\phi}'_{k}{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]){\phi}|{\theta}_k])|\mathcal{G}(k)]\\
&=&((\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k - \gamma \bm{\phi}'_{k}\bm{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k])\bm{\phi}|\bm{\theta}_k])\\ &=&((\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k - \gamma {\phi}'_{k}{\phi}^{\top}_k \textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]){\phi}|{\theta}_k])\\
& &-\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k|\bm{\theta}_k] - \gamma\mathbb{E}[\bm{\phi}' \bm{\phi}^{\top}]\textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|\bm{\theta}_k]) \bm{\phi}_k|\bm{\theta}_k]. & &-\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k|{\theta}_k] - \gamma\mathbb{E}[{\phi}' {\phi}^{\top}]\textbf{C}^{-1}\mathbb{E}[(\delta_k -\mathbb{E}[\delta_k|{\theta}_k]) {\phi}_k|{\theta}_k].
\end{array} \end{array}
\end{equation*} \end{equation*}
It is easy to see that $Z_k$, $k\geq 0$ are integrable random variables and $\mathbb{E}[Z_{k+1}|\mathcal{G}(k)]=0$, $\forall k\geq0$. Further, It is easy to see that $Z_k$, $k\geq 0$ are integrable random variables and $\mathbb{E}[Z_{k+1}|\mathcal{G}(k)]=0$, $\forall k\geq0$. Further,
\begin{equation*} \begin{equation*}
\mathbb{E}[||Z_{k+1}||^2|\mathcal{G}(k)]\leq \mathbb{E}[||Z_{k+1}||^2|\mathcal{G}(k)]\leq
c_3(1+||\bm{\theta}_k||^2), k\geq 0 c_3(1+||{\theta}_k||^2), k\geq 0
\end{equation*} \end{equation*}
for some constant $c_3 \geq 0$, again beacuse $\bm{\phi}_k$, $r_k$, and $\bm{\phi}_k'$ have for some constant $c_3 \geq 0$, again beacuse ${\phi}_k$, $r_k$, and ${\phi}_k'$ have
uniformly bounded second moments, it can be seen that for some constant. uniformly bounded second moments, it can be seen that for some constant.
Consider now the following ODE associated with (5): Consider now the following ODE associated with (5):
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t) = (\textbf{I} - \mathbb{E}[\gamma \bm{\phi}' \bm{\phi}^{\top}]\textbf{C}^{-1})\mathbb{E}[(\delta -\mathbb{E}[\delta|\bm{\theta}(t)]) \bm{\phi}|\bm{\theta}(t)]. \dot{{\theta}}(t) = (\textbf{I} - \mathbb{E}[\gamma {\phi}' {\phi}^{\top}]\textbf{C}^{-1})\mathbb{E}[(\delta -\mathbb{E}[\delta|{\theta}(t)]) {\phi}|{\theta}(t)].
\label{thetavmtdcSlowerFinal} \label{thetavmtdcSlowerFinal}
\end{equation} \end{equation}
Let Let
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\vec{h}(\bm{\theta}(t))&=&(\textbf{I} - \mathbb{E}[\gamma \bm{\phi}' \bm{\phi}^{\top}]\textbf{C}^{-1})\mathbb{E}[(\delta -\mathbb{E}[\delta|\bm{\theta}(t)]) \bm{\phi}|\bm{\theta}(t)]\\ \vec{h}({\theta}(t))&=&(\textbf{I} - \mathbb{E}[\gamma {\phi}' {\phi}^{\top}]\textbf{C}^{-1})\mathbb{E}[(\delta -\mathbb{E}[\delta|{\theta}(t)]) {\phi}|{\theta}(t)]\\
&=&(\textbf{C} - \mathbb{E}[\gamma \bm{\phi}' \bm{\phi}^{\top}])\textbf{C}^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta|\bm{\theta}(t)]) \bm{\phi}|\bm{\theta}(t)]\\ &=&(\textbf{C} - \mathbb{E}[\gamma {\phi}' {\phi}^{\top}])\textbf{C}^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta|{\theta}(t)]) {\phi}|{\theta}(t)]\\
&=& (\mathbb{E}[\bm{\phi} \bm{\phi}^{\top}] - \mathbb{E}[\gamma \bm{\phi}' \bm{\phi}^{\top}])\textbf{C}^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta|\bm{\theta}(t)]) \bm{\phi}|\bm{\theta}(t)]\\ &=& (\mathbb{E}[{\phi} {\phi}^{\top}] - \mathbb{E}[\gamma {\phi}' {\phi}^{\top}])\textbf{C}^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta|{\theta}(t)]) {\phi}|{\theta}(t)]\\
&=& \textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}\bm{\theta}(t)+\bm{b}), &=& \textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}{\theta}(t)+{b}),
\end{array} \end{array}
\end{equation*} \end{equation*}
because $\mathbb{E}[(\delta -\mathbb{E}[\delta|\bm{\theta}(t)]) \bm{\phi}|\bm{\theta}(t)]=-\textbf{A}\bm{\theta}(t)+\bm{b}$, where because $\mathbb{E}[(\delta -\mathbb{E}[\delta|{\theta}(t)]) {\phi}|{\theta}(t)]=-\textbf{A}{\theta}(t)+{b}$, where
$\textbf{A} = \mathrm{Cov}(\bm{\phi},\bm{\phi}-\gamma\bm{\phi}')$, $\bm{b}=\mathrm{Cov}(r,\bm{\phi})$, and $\textbf{C}=\mathbb{E}[\bm{\phi}\bm{\phi}^{\top}]$ $\textbf{A} = \mathrm{Cov}({\phi},{\phi}-\gamma{\phi}')$, ${b}=\mathrm{Cov}(r,{\phi})$, and $\textbf{C}=\mathbb{E}[{\phi}{\phi}^{\top}]$
Therefore, Therefore,
$\bm{\theta}^*=\textbf{A}^{-1}\bm{b}$ can be seen to be the unique globally asymptotically ${\theta}^*=\textbf{A}^{-1}{b}$ can be seen to be the unique globally asymptotically
stable equilibrium for ODE (\ref{thetavmtdcSlowerFinal}). stable equilibrium for ODE (\ref{thetavmtdcSlowerFinal}).
Let $\vec{h}_{\infty}(\bm{\theta})=\lim_{r\rightarrow Let $\vec{h}_{\infty}({\theta})=\lim_{r\rightarrow
\infty}\frac{\vec{h}(r\bm{\theta})}{r}$. Then \infty}\frac{\vec{h}(r{\theta})}{r}$. Then
$\vec{h}_{\infty}(\bm{\theta})=-\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}\bm{\theta}$ is well-defined. $\vec{h}_{\infty}({\theta})=-\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}{\theta}$ is well-defined.
Consider now the ODE Consider now the ODE
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t)=-\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}\bm{\theta}(t). \dot{{\theta}}(t)=-\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}{\theta}(t).
\label{odethetavmtdcfinal} \label{odethetavmtdcfinal}
\end{equation} \end{equation}
...@@ -462,11 +627,11 @@ Thus, the assumption (A1) and (A2) are verified. ...@@ -462,11 +627,11 @@ Thus, the assumption (A1) and (A2) are verified.
The proof is given above. The proof is given above.
In the fastest time scale, the parameter $w$ converges to In the fastest time scale, the parameter $w$ converges to
$\mathbb{E}[\delta|\bm{u}_k,\bm{\theta}_k]$. $\mathbb{E}[\delta|{u}_k,{\theta}_k]$.
In the second fast time scale, In the second fast time scale,
the parameter $u$ converges to $\textbf{C}^{-1}\mathbb{E}[(\delta-\mathbb{E}[\delta|\bm{\theta}_k])\bm{\phi}|\bm{\theta}_k]$. the parameter $u$ converges to $\textbf{C}^{-1}\mathbb{E}[(\delta-\mathbb{E}[\delta|{\theta}_k]){\phi}|{\theta}_k]$.
In the slower time scale, In the slower time scale,
the parameter $\bm{\theta}$ converges to $\textbf{A}^{-1}\bm{b}$. the parameter ${\theta}$ converges to $\textbf{A}^{-1}{b}$.
\end{proof} \end{proof}
\subsection{Proof of Theorem 2} \subsection{Proof of Theorem 2}
...@@ -478,49 +643,49 @@ the parameter $\bm{\theta}$ converges to $\textbf{A}^{-1}\bm{b}$. ...@@ -478,49 +643,49 @@ the parameter $\bm{\theta}$ converges to $\textbf{A}^{-1}\bm{b}$.
\cite{borkar1997stochastic}. \cite{borkar1997stochastic}.
The VMTD's solution is The VMTD's solution is
$\bm{\theta}_{\text{VMETD}}=\textbf{A}_{\text{VMETD}}^{-1}\bm{b}_{\text{VMETD}}$. ${\theta}_{\text{VMETD}}=\textbf{A}_{\text{VMETD}}^{-1}{b}_{\text{VMETD}}$.
First, note that recursion (19) can be rewritten as First, note that recursion (19) can be rewritten as
\begin{equation*} \begin{equation*}
\bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\beta_k\bm{\xi}(k), {\theta}_{k+1}\leftarrow {\theta}_k+\beta_k{\xi}(k),
\end{equation*} \end{equation*}
where where
\begin{equation*} \begin{equation*}
\bm{\xi}(k)=\frac{\alpha_k}{\beta_k} (F_k \rho_k\delta_k - \omega_{k+1})\bm{\phi}_k {\xi}(k)=\frac{\alpha_k}{\beta_k} (F_k \rho_k\delta_k - \omega_{k+1}){\phi}_k
\end{equation*} \end{equation*}
Due to the settings of step-size schedule $\alpha_k = o(\beta_k)$, Due to the settings of step-size schedule $\alpha_k = o(\beta_k)$,
$\bm{\xi}(k)\rightarrow 0$ almost surely as $k\rightarrow\infty$. ${\xi}(k)\rightarrow 0$ almost surely as $k\rightarrow\infty$.
That is the increments in iteration (13) are uniformly larger than That is the increments in iteration (13) are uniformly larger than
those in (12), thus (13) is the faster recursion. those in (12), thus (13) is the faster recursion.
Along the faster time scale, iterations of (12) and (13) Along the faster time scale, iterations of (12) and (13)
are associated to ODEs system as follows: are associated to ODEs system as follows:
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t) = 0, \dot{{\theta}}(t) = 0,
\label{thetaFast} \label{thetaFast}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\dot{\omega}(t)=\mathbb{E}_{\mu}[F_t\rho_t\delta_t|\bm{\theta}(t)]-\omega(t). \dot{\omega}(t)=\mathbb{E}_{\mu}[F_t\rho_t\delta_t|{\theta}(t)]-\omega(t).
\label{omegaFast} \label{omegaFast}
\end{equation} \end{equation}
Based on the ODE (\ref{thetaFast}), $\bm{\theta}(t)\equiv \bm{\theta}$ when Based on the ODE (\ref{thetaFast}), ${\theta}(t)\equiv {\theta}$ when
viewed from the faster timescale. viewed from the faster timescale.
By the Hirsch lemma \cite{hirsch1989convergent}, it follows that By the Hirsch lemma \cite{hirsch1989convergent}, it follows that
$||\bm{\theta}_k-\bm{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some $||{\theta}_k-{\theta}||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
$\bm{\theta}$ that depends on the initial condition $\bm{\theta}_0$ of recursion ${\theta}$ that depends on the initial condition ${\theta}_0$ of recursion
(12). (12).
Thus, the ODE pair (\ref{thetaFast})-(\ref{omegaFast}) can be written as Thus, the ODE pair (\ref{thetaFast})-(\ref{omegaFast}) can be written as
\begin{equation} \begin{equation}
\dot{\omega}(t)=\mathbb{E}_{\mu}[F_t\rho_t\delta_t|\bm{\theta}]-\omega(t). \dot{\omega}(t)=\mathbb{E}_{\mu}[F_t\rho_t\delta_t|{\theta}]-\omega(t).
\label{omegaFastFinal} \label{omegaFastFinal}
\end{equation} \end{equation}
Consider the function $h(\omega)=\mathbb{E}_{\mu}[F\rho\delta|\bm{\theta}]-\omega$, Consider the function $h(\omega)=\mathbb{E}_{\mu}[F\rho\delta|{\theta}]-\omega$,
i.e., the driving vector field of the ODE (\ref{omegaFastFinal}). i.e., the driving vector field of the ODE (\ref{omegaFastFinal}).
It is easy to find that the function $h$ is Lipschitz with coefficient It is easy to find that the function $h$ is Lipschitz with coefficient
$-1$. $-1$.
Let $h_{\infty}(\cdot)$ be the function defined by Let $h_{\infty}(\cdot)$ be the function defined by
$h_{\infty}(\omega)=\lim_{x\rightarrow \infty}\frac{h(x\omega)}{x}$. $h_{\infty}(\omega)=\lim_{x\rightarrow \infty}\frac{h(x\omega)}{x}$.
Then $h_{\infty}(\omega)= -\omega$, is well-defined. Then $h_{\infty}(\omega)= -\omega$, is well-defined.
For (\ref{omegaFastFinal}), $\omega^*=\mathbb{E}_{\mu}[F\rho\delta|\bm{\theta}]$ For (\ref{omegaFastFinal}), $\omega^*=\mathbb{E}_{\mu}[F\rho\delta|{\theta}]$
is the unique globally asymptotically stable equilibrium. is the unique globally asymptotically stable equilibrium.
For the ODE For the ODE
\begin{equation} \begin{equation}
...@@ -537,18 +702,18 @@ The VMTD's solution is ...@@ -537,18 +702,18 @@ The VMTD's solution is
Let Let
$M_{k+1}=(F_k\rho_k\delta_k-\omega_k) $M_{k+1}=(F_k\rho_k\delta_k-\omega_k)
-\mathbb{E}_{\mu}[(F_k\rho_k\delta_k-\omega_k)|\mathcal{F}(k)]$, -\mathbb{E}_{\mu}[(F_k\rho_k\delta_k-\omega_k)|\mathcal{F}(k)]$,
where $\mathcal{F}(k)=\sigma(\omega_l,\bm{\theta}_l,l\leq k;\bm{\phi}_s,\bm{\phi}_s',r_s,s<k)$, where $\mathcal{F}(k)=\sigma(\omega_l,{\theta}_l,l\leq k;{\phi}_s,{\phi}_s',r_s,s<k)$,
$k\geq 1$ are the sigma fields $k\geq 1$ are the sigma fields
generated by $\omega_0,\bm{\theta}_0,\omega_{l+1},\bm{\theta}_{l+1},\bm{\phi}_l,\bm{\phi}_l'$, generated by $\omega_0,{\theta}_0,\omega_{l+1},{\theta}_{l+1},{\phi}_l,{\phi}_l'$,
$0\leq l<k$. $0\leq l<k$.
It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$. satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$.
Because $\bm{\phi}_k$, $r_k$, and $\bm{\phi}_k'$ have Because ${\phi}_k$, $r_k$, and ${\phi}_k'$ have
uniformly bounded second moments, it can be seen that for some constant uniformly bounded second moments, it can be seen that for some constant
$c_1>0$, $\forall k\geq0$, $c_1>0$, $\forall k\geq0$,
\begin{equation*} \begin{equation*}
\mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq \mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq
c_1(1+||\omega_k||^2+||\bm{\theta}_k||^2). c_1(1+||\omega_k||^2+||{\theta}_k||^2).
\end{equation*} \end{equation*}
...@@ -561,23 +726,23 @@ The VMTD's solution is ...@@ -561,23 +726,23 @@ The VMTD's solution is
Consider now the slower time scale recursion (12). Consider now the slower time scale recursion (12).
Based on the above analysis, (12) can be rewritten as Based on the above analysis, (12) can be rewritten as
% \begin{equation*} % \begin{equation*}
% \bm{\theta}_{k+1}\leftarrow % {\theta}_{k+1}\leftarrow
% \bm{\theta}_{k}+\alpha_k(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k])\bm{\phi}_k. % {\theta}_{k}+\alpha_k(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]){\phi}_k.
% \end{equation*} % \end{equation*}
\begin{equation*} \begin{equation*}
\begin{split} \begin{split}
\bm{\theta}_{k+1}&\leftarrow \bm{\theta}_k+\alpha_k (F_k \rho_k\delta_k - \omega_k)\bm{\phi}_k -\alpha_k \omega_{k+1}\bm{\phi}_k\\ {\theta}_{k+1}&\leftarrow {\theta}_k+\alpha_k (F_k \rho_k\delta_k - \omega_k){\phi}_k -\alpha_k \omega_{k+1}{\phi}_k\\
&=\bm{\theta}_{k}+\alpha_k(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k])\bm{\phi}_k\\ &={\theta}_{k}+\alpha_k(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]){\phi}_k\\
&=\bm{\theta}_k+\alpha_k F_k \rho_k (R_{k+1}+\gamma \bm{\theta}_k^{\top}\bm{\phi}_{k+1}-\bm{\theta}_k^{\top}\bm{\phi}_k)\bm{\phi}_k -\alpha_k \mathbb{E}_{\mu}[F_k \rho_k \delta_k]\bm{\phi}_k\\ &={\theta}_k+\alpha_k F_k \rho_k (R_{k+1}+\gamma {\theta}_k^{\top}{\phi}_{k+1}-{\theta}_k^{\top}{\phi}_k){\phi}_k -\alpha_k \mathbb{E}_{\mu}[F_k \rho_k \delta_k]{\phi}_k\\
&= \bm{\theta}_k+\alpha_k \{\underbrace{(F_k\rho_kR_{k+1}-\mathbb{E}_{\mu}[F_k\rho_k R_{k+1}])\bm{\phi}_k}_{\bm{b}_{\text{VMETD},k}} &= {\theta}_k+\alpha_k \{\underbrace{(F_k\rho_kR_{k+1}-\mathbb{E}_{\mu}[F_k\rho_k R_{k+1}]){\phi}_k}_{{b}_{\text{VMETD},k}}
-\underbrace{(F_k\rho_k\bm{\phi}_k(\bm{\phi}_k-\gamma\bm{\phi}_{k+1})^{\top}-\bm{\phi}_k\mathbb{E}_{\mu}[F_k\rho_k (\bm{\phi}_k-\gamma\bm{\phi}_{k+1})]^{\top})}_{\textbf{A}_{\text{VMETD},k}}\bm{\theta}_k\} -\underbrace{(F_k\rho_k{\phi}_k({\phi}_k-\gamma{\phi}_{k+1})^{\top}-{\phi}_k\mathbb{E}_{\mu}[F_k\rho_k ({\phi}_k-\gamma{\phi}_{k+1})]^{\top})}_{\textbf{A}_{\text{VMETD},k}}{\theta}_k\}
\end{split} \end{split}
\end{equation*} \end{equation*}
Let $\mathcal{G}(k)=\sigma(\bm{\theta}_l,l\leq k;\bm{\phi}_s,\bm{\phi}_s',r_s,s<k)$, Let $\mathcal{G}(k)=\sigma({\theta}_l,l\leq k;{\phi}_s,{\phi}_s',r_s,s<k)$,
$k\geq 1$ be the sigma fields $k\geq 1$ be the sigma fields
generated by $\bm{\theta}_0,\bm{\theta}_{l+1},\bm{\phi}_l,\bm{\phi}_l'$, generated by ${\theta}_0,{\theta}_{l+1},{\phi}_l,{\phi}_l'$,
$0\leq l<k$. $0\leq l<k$.
Let Let
$ $
...@@ -585,17 +750,17 @@ The VMTD's solution is ...@@ -585,17 +750,17 @@ The VMTD's solution is
$ $
where where
\begin{equation*} \begin{equation*}
Y_{k}=(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k])\bm{\phi}_k. Y_{k}=(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]){\phi}_k.
\end{equation*} \end{equation*}
Consequently, Consequently,
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\mathbb{E}_{\mu}[Y_k|\mathcal{G}(k)]&=&\mathbb{E}_{\mu}[(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k])\bm{\phi}_k|\mathcal{G}(k)]\\ \mathbb{E}_{\mu}[Y_k|\mathcal{G}(k)]&=&\mathbb{E}_{\mu}[(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]){\phi}_k|\mathcal{G}(k)]\\
&=&\mathbb{E}_{\mu}[F_k\rho_k\delta_k\bm{\phi}_k|\bm{\theta}_k] &=&\mathbb{E}_{\mu}[F_k\rho_k\delta_k{\phi}_k|{\theta}_k]
-\mathbb{E}_{\mu}[\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k]\bm{\phi}_k]\\ -\mathbb{E}_{\mu}[\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]{\phi}_k]\\
&=&\mathbb{E}_{\mu}[F_k\rho_k\delta_k\bm{\phi}_k|\bm{\theta}_k] &=&\mathbb{E}_{\mu}[F_k\rho_k\delta_k{\phi}_k|{\theta}_k]
-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k]\mathbb{E}_{\mu}[\bm{\phi}_k]\\ -\mathbb{E}_{\mu}[F_k\rho_k\delta_k|{\theta}_k]\mathbb{E}_{\mu}[{\phi}_k]\\
&=&\mathrm{Cov}(F_k\rho_k\delta_k|\bm{\theta}_k,\bm{\phi}_k), &=&\mathrm{Cov}(F_k\rho_k\delta_k|{\theta}_k,{\phi}_k),
\end{array} \end{array}
\end{equation*} \end{equation*}
where $\mathrm{Cov}(\cdot,\cdot)$ is a covariance operator. where $\mathrm{Cov}(\cdot,\cdot)$ is a covariance operator.
...@@ -603,61 +768,61 @@ The VMTD's solution is ...@@ -603,61 +768,61 @@ The VMTD's solution is
Thus, Thus,
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
Z_{k+1}&=&(F_k\rho_k\delta_k-\mathbb{E}[\delta_k|\bm{\theta}_k])\bm{\phi}_k-\mathrm{Cov}(F_k\rho_k\delta_k|\bm{\theta}_k,\bm{\phi}_k). Z_{k+1}&=&(F_k\rho_k\delta_k-\mathbb{E}[\delta_k|{\theta}_k]){\phi}_k-\mathrm{Cov}(F_k\rho_k\delta_k|{\theta}_k,{\phi}_k).
\end{array} \end{array}
\end{equation*} \end{equation*}
It is easy to verify that $Z_{k+1},k\geq0$ are integrable random variables that It is easy to verify that $Z_{k+1},k\geq0$ are integrable random variables that
satisfy $\mathbb{E}[Z_{k+1}|\mathcal{G}(k)]=0$, $\forall k\geq0$. satisfy $\mathbb{E}[Z_{k+1}|\mathcal{G}(k)]=0$, $\forall k\geq0$.
Also, because $\bm{\phi}_k$, $r_k$, and $\bm{\phi}_k'$ have Also, because ${\phi}_k$, $r_k$, and ${\phi}_k'$ have
uniformly bounded second moments, it can be seen that for some constant uniformly bounded second moments, it can be seen that for some constant
$c_2>0$, $\forall k\geq0$, $c_2>0$, $\forall k\geq0$,
\begin{equation*} \begin{equation*}
\mathbb{E}[||Z_{k+1}||^2|\mathcal{G}(k)]\leq \mathbb{E}[||Z_{k+1}||^2|\mathcal{G}(k)]\leq
c_2(1+||\bm{\theta}_k||^2). c_2(1+||{\theta}_k||^2).
\end{equation*} \end{equation*}
Consider now the following ODE associated with (12): Consider now the following ODE associated with (12):
\begin{equation} \begin{equation}
\begin{array}{ccl} \begin{array}{ccl}
\dot{\bm{\theta}}(t)&=&-\textbf{A}_{\text{VMETD}}\bm{\theta}(t)+\bm{b}_{\text{VMETD}}. \dot{{\theta}}(t)&=&-\textbf{A}_{\text{VMETD}}{\theta}(t)+{b}_{\text{VMETD}}.
\end{array} \end{array}
\label{odetheta} \label{odetheta}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\begin{split} \begin{split}
\textbf{A}_{\text{VMETD}}&=\lim_{k \rightarrow \infty} \mathbb{E}[\textbf{A}_{\text{VMETD},k}]\\ \textbf{A}_{\text{VMETD}}&=\lim_{k \rightarrow \infty} \mathbb{E}[\textbf{A}_{\text{VMETD},k}]\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k \rho_k \bm{\phi}_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k \rho_k {\phi}_k ({\phi}_k - \gamma {\phi}_{k+1})^{\top}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})]^{\top}\\
% &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\underbrace{\bm{\phi}_k}_{X}\underbrace{F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}}_{Y}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ % &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\underbrace{{\phi}_k}_{X}\underbrace{F_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})^{\top}}_{Y}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})]^{\top}\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_kF_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})^{\top}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})]^{\top}\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]- \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_kF_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})^{\top}]- \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k \rho_k ({\phi}_k - \gamma {\phi}_{k+1})]^{\top}\\
&=\sum_{s} f(s) \bm{\phi}(s)(\bm{\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\bm{\phi}(s'))^{\top} - \sum_{s} d_{\mu}(s) \bm{\phi}(s) * \sum_{s} f(s)(\bm{\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\bm{\phi}(s'))^{\top} \\ &=\sum_{s} f(s) {\phi}(s)({\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}{\phi}(s'))^{\top} - \sum_{s} d_{\mu}(s) {\phi}(s) * \sum_{s} f(s)({\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}{\phi}(s'))^{\top} \\
&={\bm{\Phi}}^{\top} \textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}) \bm{\Phi} - {\bm{\Phi}}^{\top} \textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\mu}) \bm{\Phi} \\ &={{\Phi}}^{\top} \textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}) {\Phi} - {{\Phi}}^{\top} \textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\mu}) {\Phi} \\
&={\bm{\Phi}}^{\top} (\textbf{F} - \textbf{d}_{\mu} \textbf{f}^{\top}) (\textbf{I} - \gamma \textbf{P}_{\pi}){\bm{\Phi}} \\ &={{\Phi}}^{\top} (\textbf{F} - \textbf{d}_{\mu} \textbf{f}^{\top}) (\textbf{I} - \gamma \textbf{P}_{\pi}){{\Phi}} \\
&={\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\pi})){\bm{\Phi}} \\ &={{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\pi})){{\Phi}} \\
&={\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){\bm{\Phi}} \\ &={{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){{\Phi}} \\
\end{split} \end{split}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\begin{split} \begin{split}
\bm{b}_{\text{VMETD}}&=\lim_{k \rightarrow \infty} \mathbb{E}[\bm{b}_{\text{VMETD},k}]\\ {b}_{\text{VMETD}}&=\lim_{k \rightarrow \infty} \mathbb{E}[{b}_{\text{VMETD},k}]\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k\rho_kR_{k+1}\bm{\phi}_k]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k\rho_kR_{k+1}{\phi}_k]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k\rho_kR_{k+1}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[\bm{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_kF_k\rho_kR_{k+1}]- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\mathbb{E}_{\mu}[{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\
&= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k\rho_kR_{k+1}]- \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &= \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_kF_k\rho_kR_{k+1}]- \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\
&=\sum_{s} f(s) \bm{\phi}(s)r_{\pi} - \sum_{s} d_{\mu}(s) \bm{\phi}(s) * \sum_{s} f(s)r_{\pi} \\ &=\sum_{s} f(s) {\phi}(s)r_{\pi} - \sum_{s} d_{\mu}(s) {\phi}(s) * \sum_{s} f(s)r_{\pi} \\
&=\bm{\bm{\Phi}}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi} \\ &={{\Phi}}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi} \\
\end{split} \end{split}
\end{equation} \end{equation}
Let $\vec{h}(\bm{\theta}(t))$ be the driving vector field of the ODE Let $\vec{h}({\theta}(t))$ be the driving vector field of the ODE
(\ref{odetheta}). (\ref{odetheta}).
\begin{equation*} \begin{equation*}
\vec{h}(\bm{\theta}(t))=-\textbf{A}_{\text{VMETD}}\bm{\theta}(t)+\bm{b}_{\text{VMETD}}. \vec{h}({\theta}(t))=-\textbf{A}_{\text{VMETD}}{\theta}(t)+{b}_{\text{VMETD}}.
\end{equation*} \end{equation*}
An $\bm{\Phi}^{\top}\bm{\text{X}}\bm{\Phi}$ matrix of this An ${\Phi}^{\top}{\text{X}}{\Phi}$ matrix of this
form will be positive definite whenever the matrix $\bm{\text{X}}$ is positive definite. form will be positive definite whenever the matrix ${\text{X}}$ is positive definite.
Any matrix $\bm{\text{X}}$ is positive definite if and only if Any matrix ${\text{X}}$ is positive definite if and only if
the symmetric matrix $\bm{\text{S}}=\bm{\text{X}}+\bm{\text{X}}^{\top}$ is positive definite. the symmetric matrix ${\text{S}}={\text{X}}+{\text{X}}^{\top}$ is positive definite.
Any symmetric real matrix $\bm{\text{S}}$ is positive definite if the absolute values of Any symmetric real matrix ${\text{S}}$ is positive definite if the absolute values of
its diagonal entries are greater than the sum of the absolute values of the corresponding its diagonal entries are greater than the sum of the absolute values of the corresponding
off-diagonal entries\cite{sutton2016emphatic}. off-diagonal entries\cite{sutton2016emphatic}.
...@@ -693,14 +858,14 @@ The VMTD's solution is ...@@ -693,14 +858,14 @@ The VMTD's solution is
Therefore, Therefore,
$\bm{\theta}^*=\textbf{A}_{\text{VMETD}}^{-1}\bm{b}_{\text{VMETD}}$ can be seen to be the unique globally asymptotically ${\theta}^*=\textbf{A}_{\text{VMETD}}^{-1}{b}_{\text{VMETD}}$ can be seen to be the unique globally asymptotically
stable equilibrium for ODE (\ref{odetheta}). stable equilibrium for ODE (\ref{odetheta}).
Let $\vec{h}_{\infty}(\bm{\theta})=\lim_{r\rightarrow Let $\vec{h}_{\infty}({\theta})=\lim_{r\rightarrow
\infty}\frac{\vec{h}(r\bm{\theta})}{r}$. Then \infty}\frac{\vec{h}(r{\theta})}{r}$. Then
$\vec{h}_{\infty}(\bm{\theta})=-\textbf{A}_{\text{VMETD}}\bm{\theta}$ is well-defined. $\vec{h}_{\infty}({\theta})=-\textbf{A}_{\text{VMETD}}{\theta}$ is well-defined.
Consider now the ODE Consider now the ODE
\begin{equation} \begin{equation}
\dot{\bm{\theta}}(t)=-\textbf{A}_{\text{VMETD}}\bm{\theta}(t). \dot{{\theta}}(t)=-\textbf{A}_{\text{VMETD}}{\theta}(t).
\label{odethetafinal} \label{odethetafinal}
\end{equation} \end{equation}
The ODE (\ref{odethetafinal}) has the origin as its unique globally asymptotically stable equilibrium. The ODE (\ref{odethetafinal}) has the origin as its unique globally asymptotically stable equilibrium.
...@@ -719,25 +884,25 @@ The VMTD's solution is ...@@ -719,25 +884,25 @@ The VMTD's solution is
% { % {
% \begin{tabular}{cccc} % \begin{tabular}{cccc}
% \toprule % \toprule
% Algorithm&Key matrix $\textbf{A}$&{Positive definite}&{$\bm{b}$}\\\midrule % Algorithm&Key matrix $\textbf{A}$&{Positive definite}&{${b}$}\\\midrule
% On-policy TD&$\bm{\Phi}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma % On-policy TD&${\Phi}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma
% \textbf{P}_{\pi})\bm{\Phi}$&$\checkmark$&$\bm{b}_{\text{on}}=\bm{\Phi}^{\top}\textbf{D}_{\pi}\textbf{r}_{\pi}$\\ % \textbf{P}_{\pi}){\Phi}$&$\checkmark$&${b}_{\text{on}}={\Phi}^{\top}\textbf{D}_{\pi}\textbf{r}_{\pi}$\\
% On-policy VMTD&${\bm{\Phi}}^{\top}(\textbf{D}_{\pi}-\textbf{d}_{\pi} \textbf{d}_{\pi}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){\bm{\Phi}}$ % On-policy VMTD&${{\Phi}}^{\top}(\textbf{D}_{\pi}-\textbf{d}_{\pi} \textbf{d}_{\pi}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){{\Phi}}$
% &$\checkmark$&$\bm{\Phi}^{\top}(\textbf{D}_{\pi}-\textbf{d}_{\pi} \textbf{d}_{\pi}^{\top})\textbf{r}_{\pi}$\\ % &$\checkmark$&${\Phi}^{\top}(\textbf{D}_{\pi}-\textbf{d}_{\pi} \textbf{d}_{\pi}^{\top})\textbf{r}_{\pi}$\\
% \midrule % \midrule
% Off-policy TD&$\textbf{A}_{\text{off}}={\bm{\Phi}}^{\top}\textbf{D}_{\mu}(\textbf{I}-\gamma % Off-policy TD&$\textbf{A}_{\text{off}}={{\Phi}}^{\top}\textbf{D}_{\mu}(\textbf{I}-\gamma
% \textbf{P}_{\pi}){\bm{\Phi}}$&$\times$&$\bm{b}_{\text{off}}=\bm{\Phi}^{\top}\textbf{D}_{\mu}\textbf{r}_{\pi}$\\ % \textbf{P}_{\pi}){{\Phi}}$&$\times$&${b}_{\text{off}}={\Phi}^{\top}\textbf{D}_{\mu}\textbf{r}_{\pi}$\\
% TDC& $\textbf{A}_{\text{off}}^{\top}\textbf{C}^{-1}\textbf{A}_{\text{off}}$&$\checkmark$&$\textbf{A}_{\text{off}}^{\top}\textbf{C}^{-1}\bm{b}_{\text{off}}$ % TDC& $\textbf{A}_{\text{off}}^{\top}\textbf{C}^{-1}\textbf{A}_{\text{off}}$&$\checkmark$&$\textbf{A}_{\text{off}}^{\top}\textbf{C}^{-1}{b}_{\text{off}}$
% \\ % \\
% ETD& ${\bm{\Phi}}^{\top}\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}){\bm{\Phi}}$ % ETD& ${{\Phi}}^{\top}\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}){{\Phi}}$
% &$\checkmark$&$\bm{\Phi}^{\top}\textbf{F}\textbf{r}_{\pi}$\\ % &$\checkmark$&${\Phi}^{\top}\textbf{F}\textbf{r}_{\pi}$\\
% \midrule % \midrule
% Off-policy VMTD&$\textbf{A}_{\text{VMTD}}={\bm{\Phi}}^{\top} (\textbf{D}_{\mu}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){\bm{\Phi}}$ % Off-policy VMTD&$\textbf{A}_{\text{VMTD}}={{\Phi}}^{\top} (\textbf{D}_{\mu}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} )(\textbf{I} - \gamma\textbf{P}_{\pi}){{\Phi}}$
% &$\times$&$\bm{b}_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top})\textbf{r}_{\pi}$\\ % &$\times$&${b}_{\text{VMTD}}={\Phi}^{\top}(\textbf{D}_{\mu}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top})\textbf{r}_{\pi}$\\
% VMTDC& $\textbf{A}_{\text{VMTD}}^{\top}\textbf{C}^{-1}\textbf{A}_{\text{VMTD}}$&$\checkmark$&$\textbf{A}_{\text{VMTD}}^{\top}\textbf{C}^{-1}\bm{b}_{\text{VMTD}}$ % VMTDC& $\textbf{A}_{\text{VMTD}}^{\top}\textbf{C}^{-1}\textbf{A}_{\text{VMTD}}$&$\checkmark$&$\textbf{A}_{\text{VMTD}}^{\top}\textbf{C}^{-1}{b}_{\text{VMTD}}$
% \\ % \\
% VMETD& ${\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){\bm{\Phi}}$ % VMETD& ${{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){{\Phi}}$
% &$\checkmark$&$\bm{\Phi}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi}$\\ % &$\checkmark$&${\Phi}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi}$\\
% \bottomrule % \bottomrule
% \end{tabular} % \end{tabular}
% } % }
...@@ -835,50 +1000,57 @@ The VMTD's solution is ...@@ -835,50 +1000,57 @@ The VMTD's solution is
\section{Experimental details} \section{Experimental details}
\label{experimentaldetails} \label{experimentaldetails}
The 2-state counterexample and the 7-state counterexample % The 2-state counterexample and the 7-state counterexample
are well-known off-policy experimental environments. The 2-state % are well-known off-policy experimental environments. The 2-state
counterexample is relatively simple, so next, I'll provide a detailed % counterexample is relatively simple, so next, I'll provide a detailed
description of the 7-state counterexample environment. % description of the 7-state counterexample environment.
\textbf{Baird's off-policy counterexample:} This task is well known as a % \textbf{Baird's off-policy counterexample:} This task is well known as a
counterexample, in which TD diverges \cite{baird1995residual,sutton2009fast}. As % counterexample, in which TD diverges \cite{baird1995residual,sutton2009fast}. As
shown in Figure \ref{bairdexample}, reward for each transition is zero. Thus the true values are zeros for all states and for any given policy. The behaviour policy % shown in Figure \ref{bairdexample}, reward for each transition is zero. Thus the true values are zeros for all states and for any given policy. The behaviour policy
chooses actions represented by solid lines with a probability of $\frac{1}{7}$ % chooses actions represented by solid lines with a probability of $\frac{1}{7}$
and actions represented by dotted lines with a probability of $\frac{6}{7}$. The % and actions represented by dotted lines with a probability of $\frac{6}{7}$. The
target policy is expected to choose the solid line with more probability than $\frac{1}{7}$, % target policy is expected to choose the solid line with more probability than $\frac{1}{7}$,
and it chooses the solid line with probability of $1$ in this paper. % and it chooses the solid line with probability of $1$ in this paper.
The discount factor $\gamma =0.99$, and the feature matrix is % The discount factor $\gamma =0.99$, and the feature matrix is
defined in Appendix \ref{experimentaldetails} \cite{baird1995residual,sutton2009fast,maei2011gradient}. % defined in Appendix \ref{experimentaldetails} \cite{baird1995residual,sutton2009fast,maei2011gradient}.
\begin{figure} % \begin{figure}
\begin{center} % \begin{center}
\input{pic/BairdExample.tex} % \input{pic/BairdExample.tex}
\caption{7-state version of Baird's off-policy counterexample.} % \caption{7-state version of Baird's off-policy counterexample.}
\label{bairdexample} % \label{bairdexample}
\end{center} % \end{center}
\end{figure} % \end{figure}
The feature matrix of 7-state version of Baird's off-policy counterexample is % The feature matrix of 7-state version of Baird's off-policy counterexample is
defined as follow: % defined as follow:
\begin{equation*} % \begin{equation*}
\Phi_{Counter}=\left[ % \Phi_{Counter}=\left[
\begin{array}{cccccccc} % \begin{array}{cccccccc}
1 & 2& 0& 0& 0& 0& 0& 0\\ % 1 & 2& 0& 0& 0& 0& 0& 0\\
1 & 0& 2& 0& 0& 0& 0& 0\\ % 1 & 0& 2& 0& 0& 0& 0& 0\\
1 & 0& 0& 2& 0& 0& 0& 0\\ % 1 & 0& 0& 2& 0& 0& 0& 0\\
1 & 0& 0& 0& 2& 0& 0& 0\\ % 1 & 0& 0& 0& 2& 0& 0& 0\\
1 & 0& 0& 0& 0& 2& 0& 0\\ % 1 & 0& 0& 0& 0& 2& 0& 0\\
1 & 0& 0& 0& 0& 0& 2& 0\\ % 1 & 0& 0& 0& 0& 0& 2& 0\\
2 & 0& 0& 0& 0& 0& 0& 1 % 2 & 0& 0& 0& 0& 0& 0& 1
\end{array}\right] % \end{array}\right]
\end{equation*} % \end{equation*}
2-state version of Baird's off-policy counterexample: All learning rates follow linear learning rate decay.
For TD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
For TDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$ and $\alpha_0 = 0.1$.
For VMTDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$, $\frac{\alpha_k}{\omega_k}=4$,and $\alpha_0 = 0.1$.
For VMTD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
2-state version of Baird's off-policy counterexample: All learning rates follow linear learning rate decay. 2-state version of Baird's off-policy counterexample: All learning rates follow linear learning rate decay.
For TD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
For TDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$ and $\alpha_0 = 0.1$.For ETD algorithm, $\alpha_0 = 0.1$. For TDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$ and $\alpha_0 = 0.1$.For ETD algorithm, $\alpha_0 = 0.1$.
For VMTDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$, $\frac{\alpha_k}{\omega_k}=4$,and $\alpha_0 = 0.1$.For ETD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$. For VMTDC algorithm, $\frac{\alpha_k}{\zeta_k}=5$, $\frac{\alpha_k}{\omega_k}=4$,and $\alpha_0 = 0.1$.For VMETD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
For VMTD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
7-state version of Baird's off-policy counterexample: All learning rates follow linear learning rate decay. % 7-state version of Baird's off-policy counterexample: All learning rates follow linear learning rate decay.
For TDC algorithm, $\frac{\alpha_k}{\zeta_k}=3$ and $\alpha_0 = 0.1$.For ETD algorithm, $\alpha_0 = 0.1$. % For TDC algorithm, $\frac{\alpha_k}{\zeta_k}=3$ and $\alpha_0 = 0.1$.For ETD algorithm, $\alpha_0 = 0.1$.
For VMTDC algorithm, $\frac{\alpha_k}{\zeta_k}=3$, $\frac{\alpha_k}{\omega_k}=4$,and $\alpha_0 = 0.1$.For ETD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$. % For VMTDC algorithm, $\frac{\alpha_k}{\zeta_k}=3$, $\frac{\alpha_k}{\omega_k}=4$,and $\alpha_0 = 0.1$.For ETD algorithm, $\frac{\alpha_k}{\omega_k}=4$ and $\alpha_0 = 0.1$.
For all policy evaluation experiments, each experiment For all policy evaluation experiments, each experiment
is independently run 100 times. is independently run 100 times.
...@@ -887,28 +1059,28 @@ For the four control experiments: The learning rates for each ...@@ -887,28 +1059,28 @@ For the four control experiments: The learning rates for each
algorithm in all experiments are shown in Table \ref{lrofways}. algorithm in all experiments are shown in Table \ref{lrofways}.
For all control experiments, each experiment is independently run 50 times. For all control experiments, each experiment is independently run 50 times.
\textbf{Maze}: The learning agent should find a shortest path from the upper % \textbf{Maze}: The learning agent should find a shortest path from the upper
left corner to the lower right corner. % left corner to the lower right corner.
In each state, % In each state,
there are four alternative actions: $up$, $down$, $left$, and $right$, which % there are four alternative actions: $up$, $down$, $left$, and $right$, which
takes the agent deterministically to the corresponding neighbour state, % takes the agent deterministically to the corresponding neighbour state,
except when a movement is blocked by an obstacle or the edge % except when a movement is blocked by an obstacle or the edge
of the maze. Rewards are $-1$ in all transitions until the % of the maze. Rewards are $-1$ in all transitions until the
agent reaches the goal state. % agent reaches the goal state.
The discount factor $\gamma=0.99$, and states $s$ are represented by tabular % The discount factor $\gamma=0.99$, and states $s$ are represented by tabular
features.The maximum number of moves in the game is set to 1000. % features.The maximum number of moves in the game is set to 1000.
\begin{figure} % \begin{figure}
\centering % \centering
\includegraphics[scale=0.35]{pic/maze_13_13.pdf} % \includegraphics[scale=0.35]{pic/maze_13_13.pdf}
\caption{Maze.} % \caption{Maze.}
\end{figure} % \end{figure}
\textbf{The other three control environments}: Cliff Walking, Mountain Car, and Acrobot are % \textbf{The other three control environments}: Cliff Walking, Mountain Car, and Acrobot are
selected from the gym official website and correspond to the following % selected from the gym official website and correspond to the following
versions: ``CliffWalking-v0'', ``MountainCar-v0'' and ``Acrobot-v1''. % versions: ``CliffWalking-v0'', ``MountainCar-v0'' and ``Acrobot-v1''.
For specific details, please refer to the gym official website. % For specific details, please refer to the gym official website.
The maximum number of steps for the Mountain Car environment is set to 1000, % The maximum number of steps for the Mountain Car environment is set to 1000,
while the default settings are used for the other two environments. In Mountain car and Acrobot, features are generated by tile coding. % while the default settings are used for the other two environments. In Mountain car and Acrobot, features are generated by tile coding.
\begin{table*}[htb] \begin{table*}[htb]
\centering \centering
...@@ -918,15 +1090,14 @@ while the default settings are used for the other two environments. In Mountain ...@@ -918,15 +1090,14 @@ while the default settings are used for the other two environments. In Mountain
\hline \hline
\multicolumn{1}{c|}{\diagbox{algorithms($lr$)}{envs}} &Maze &Cliff walking &Mountain Car &Acrobot \\ \multicolumn{1}{c|}{\diagbox{algorithms($lr$)}{envs}} &Maze &Cliff walking &Mountain Car &Acrobot \\
\hline \hline
% Sarsa($\alpha$)&$0.1$ &$0.1$ &$0.1$ &$0.1$ \\ Sarsa($\alpha$)&$0.1$ &$0.1$ &$0.1$ &$0.1$ \\
GQ($\alpha,\zeta$)&$0.1,0.003$ &$0.1,0.004$ &$0.1,0.01$ &$0.1,0.01$ \\ GQ($\alpha,\zeta$)&$0.1,0.003$ &$0.1,0.004$ &$0.1,0.01$ &$0.1,0.01$ \\
EQ($\alpha$)&$0.006$ &$0.005$ &$0.001$ &$0.0005$ \\ EQ($\alpha$)&$0.006$ &$0.005$ &$0.001$ &$0.0005$ \\
% VMSarsa($\alpha,\beta$)&$0.1,0.001$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ \\ VMSarsa($\alpha,\beta$)&$0.1,0.001$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ \\
VMGQ($\alpha,\zeta,\beta$)&$0.1,0.001,0.001$ &$0.1,0.005,\text{1e-4}$ &$0.1,\text{5e-4},\text{1e-4}$ &$0.1,\text{5e-4},\text{1e-4}$ \\ VMGQ($\alpha,\zeta,\beta$)&$0.1,0.001,0.001$ &$0.1,0.005,\text{1e-4}$ &$0.1,\text{5e-4},\text{1e-4}$ &$0.1,\text{5e-4},\text{1e-4}$ \\
VMEQ($\alpha,\beta$)&$0.001,0.0005$ &$0.005,0.0001$ &$0.001,0.0001$ &$0.0005,0.0001$ \\ VMEQ($\alpha,\beta$)&$0.001,0.0005$ &$0.005,0.0001$ &$0.001,0.0001$ &$0.0005,0.0001$ \\
% AC($lr_{\text{actor}},lr_{\text{critic}}$)&$0.01,0.1$ &$0.01,0.01$ &$0.01,0.05$ &$0.01,0.05$ \\ Q-learning($\alpha$)&$0.1$ &$0.1$ &$0.1$ &$0.1$ \\
% Q-learning($\alpha$)&$0.1$ &$0.1$ &$0.1$ &$0.1$ \\ VMQ($\alpha,\beta$)&$0.1,0.001$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ \\
% VMQ($\alpha,\beta$)&$0.1,0.001$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ &$0.1,\text{1e-4}$ \\
\hline \hline
\end{tabular} \end{tabular}
\label{lrofways} \label{lrofways}
......
NEW_aaai/anonymous-submission-latex-2025.aux
...@@ -26,81 +26,69 @@ ...@@ -26,81 +26,69 @@
\newlabel{introduction}{{}{1}} \newlabel{introduction}{{}{1}}
\citation{Sutton2018book} \citation{Sutton2018book}
\citation{Sutton2018book} \citation{Sutton2018book}
\citation{sutton2016emphatic}
\newlabel{preliminaries}{{}{2}} \newlabel{preliminaries}{{}{2}}
\newlabel{valuefunction}{{}{2}} \newlabel{valuefunction}{{}{2}}
\newlabel{linearvaluefunction}{{1}{2}} \newlabel{linearvaluefunction}{{1}{2}}
\newlabel{thetatd_onpolicy}{{}{2}}
\newlabel{thetatd_offpolicy}{{}{2}}
\newlabel{thetatdc}{{}{3}}
\newlabel{utdc}{{}{3}}
\newlabel{fvmetd}{{2}{3}}
\newlabel{thetaetd}{{}{3}}
\providecommand*\caption@xref[2]{\@setref\relax\@undefined{#1}} \providecommand*\caption@xref[2]{\@setref\relax\@undefined{#1}}
\newlabel{alg:algorithm 2}{{1}{3}} \newlabel{tab:min_eigenvalues}{{1}{3}}
\newlabel{alg:algorithm 5}{{2}{4}} \newlabel{delta}{{3}{3}}
\newlabel{thetavmtdc}{{5}{4}} \newlabel{omega}{{4}{3}}
\newlabel{uvmtdc}{{6}{4}} \newlabel{theta}{{5}{3}}
\newlabel{omegavmtdc}{{7}{4}} \newlabel{thetavmtdc}{{8}{4}}
\newlabel{rho_VPBE}{{8}{4}} \newlabel{uvmtdc}{{9}{4}}
\newlabel{omegavmtdc}{{10}{4}}
\newlabel{fvmetd}{{11}{4}}
\newlabel{thetavmetd}{{12}{4}} \newlabel{thetavmetd}{{12}{4}}
\newlabel{omegavmetd}{{13}{4}} \newlabel{omegavmetd}{{13}{4}}
\citation{borkar1997stochastic}
\citation{sutton2009fast} \citation{sutton2009fast}
\citation{hirsch1989convergent}
\newlabel{theorem2}{{1}{5}}
\newlabel{thetavmtdcFastest}{{14}{5}}
\newlabel{uvmtdcFastest}{{15}{5}}
\newlabel{omegavmtdcFastest}{{16}{5}}
\newlabel{omegavmtdcFastestFinal}{{17}{5}}
\newlabel{omegavmtdcInfty}{{18}{5}}
\citation{borkar2000ode}
\citation{borkar2000ode}
\citation{borkar2000ode}
\citation{borkar1997stochastic} \citation{borkar1997stochastic}
\newlabel{theorem1}{{1}{5}}
\newlabel{th1proof}{{}{5}}
\newlabel{covariance}{{14}{5}}
\newlabel{theorem2}{{2}{5}}
\newlabel{theorem3}{{3}{5}}
\newlabel{rowsum}{{15}{5}}
\newlabel{columnsum}{{16}{5}}
\citation{ng1999policy} \citation{ng1999policy}
\citation{devlin2012dynamic} \citation{devlin2012dynamic}
\newlabel{theorem3}{{2}{6}}
\newlabel{rowsum}{{19}{6}}
\newlabel{example_bias}{{2}{6}} \newlabel{example_bias}{{2}{6}}
\newlabel{columnsum}{{20}{6}}
\bibdata{aaai25} \bibdata{aaai25}
\bibcite{baird1995residual}{{1}{1995}{{Baird et~al.}}{{}}} \bibcite{baird1995residual}{{1}{1995}{{Baird et~al.}}{{}}}
\newlabel{2-state}{{1(a)}{7}} \newlabel{2-state}{{3(a)}{7}}
\newlabel{sub@2-state}{{(a)}{7}} \newlabel{sub@2-state}{{(a)}{7}}
\newlabel{7-state}{{1(b)}{7}} \newlabel{7-state}{{3(b)}{7}}
\newlabel{sub@7-state}{{(b)}{7}} \newlabel{sub@7-state}{{(b)}{7}}
\newlabel{MazeFull}{{1(c)}{7}} \newlabel{MazeFull}{{3(c)}{7}}
\newlabel{sub@MazeFull}{{(c)}{7}} \newlabel{sub@MazeFull}{{(c)}{7}}
\newlabel{CliffWalkingFull}{{1(d)}{7}} \newlabel{CliffWalkingFull}{{3(d)}{7}}
\newlabel{sub@CliffWalkingFull}{{(d)}{7}} \newlabel{sub@CliffWalkingFull}{{(d)}{7}}
\newlabel{MountainCarFull}{{1(e)}{7}} \newlabel{MountainCarFull}{{3(e)}{7}}
\newlabel{sub@MountainCarFull}{{(e)}{7}} \newlabel{sub@MountainCarFull}{{(e)}{7}}
\newlabel{AcrobotFull}{{1(f)}{7}} \newlabel{AcrobotFull}{{3(f)}{7}}
\newlabel{sub@AcrobotFull}{{(f)}{7}} \newlabel{sub@AcrobotFull}{{(f)}{7}}
\newlabel{Complete_full}{{1}{7}} \newlabel{Complete_full}{{3}{7}}
\bibcite{basserrano2021logistic}{{2}{2021}{{Bas-Serrano et~al.}}{{Bas-Serrano, Curi, Krause, and Neu}}} \bibcite{basserrano2021logistic}{{2}{2021}{{Bas-Serrano et~al.}}{{Bas-Serrano, Curi, Krause, and Neu}}}
\bibcite{borkar1997stochastic}{{3}{1997}{{Borkar}}{{}}} \bibcite{borkar1997stochastic}{{3}{1997}{{Borkar}}{{}}}
\bibcite{borkar2000ode}{{4}{2000}{{Borkar and Meyn}}{{}}} \bibcite{chen2023modified}{{4}{2023}{{Chen et~al.}}{{Chen, Ma, Li, Yang, Yang, and Gao}}}
\bibcite{chen2023modified}{{5}{2023}{{Chen et~al.}}{{Chen, Ma, Li, Yang, Yang, and Gao}}} \bibcite{devlin2012dynamic}{{5}{2012}{{Devlin and Kudenko}}{{}}}
\bibcite{devlin2012dynamic}{{6}{2012}{{Devlin and Kudenko}}{{}}} \bibcite{feng2019kernel}{{6}{2019}{{Feng, Li, and Liu}}{{}}}
\bibcite{feng2019kernel}{{7}{2019}{{Feng, Li, and Liu}}{{}}} \bibcite{givchi2015quasi}{{7}{2015}{{Givchi and Palhang}}{{}}}
\bibcite{givchi2015quasi}{{8}{2015}{{Givchi and Palhang}}{{}}} \bibcite{hackman2012faster}{{8}{2012}{{Hackman}}{{}}}
\bibcite{hackman2012faster}{{9}{2012}{{Hackman}}{{}}} \bibcite{hallak2016generalized}{{9}{2016}{{Hallak et~al.}}{{Hallak, Tamar, Munos, and Mannor}}}
\bibcite{hallak2016generalized}{{10}{2016}{{Hallak et~al.}}{{Hallak, Tamar, Munos, and Mannor}}} \bibcite{johnson2013accelerating}{{10}{2013}{{Johnson and Zhang}}{{}}}
\bibcite{hirsch1989convergent}{{11}{1989}{{Hirsch}}{{}}} \bibcite{korda2015td}{{11}{2015}{{Korda and La}}{{}}}
\bibcite{johnson2013accelerating}{{12}{2013}{{Johnson and Zhang}}{{}}} \bibcite{liu2018proximal}{{12}{2018}{{Liu et~al.}}{{Liu, Gemp, Ghavamzadeh, Liu, Mahadevan, and Petrik}}}
\bibcite{korda2015td}{{13}{2015}{{Korda and La}}{{}}} \bibcite{liu2015finite}{{13}{2015}{{Liu et~al.}}{{Liu, Liu, Ghavamzadeh, Mahadevan, and Petrik}}}
\bibcite{liu2018proximal}{{14}{2018}{{Liu et~al.}}{{Liu, Gemp, Ghavamzadeh, Liu, Mahadevan, and Petrik}}} \bibcite{liu2016proximal}{{14}{2016}{{Liu et~al.}}{{Liu, Liu, Ghavamzadeh, Mahadevan, and Petrik}}}
\bibcite{liu2015finite}{{15}{2015}{{Liu et~al.}}{{Liu, Liu, Ghavamzadeh, Mahadevan, and Petrik}}} \bibcite{ng1999policy}{{15}{1999}{{Ng, Harada, and Russell}}{{}}}
\bibcite{liu2016proximal}{{16}{2016}{{Liu et~al.}}{{Liu, Liu, Ghavamzadeh, Mahadevan, and Petrik}}} \bibcite{pan2017accelerated}{{16}{2017}{{Pan, White, and White}}{{}}}
\bibcite{ng1999policy}{{17}{1999}{{Ng, Harada, and Russell}}{{}}} \bibcite{sutton2009fast}{{17}{2009}{{Sutton et~al.}}{{Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}}}
\bibcite{pan2017accelerated}{{18}{2017}{{Pan, White, and White}}{{}}} \bibcite{sutton1988learning}{{18}{1988}{{Sutton}}{{}}}
\bibcite{sutton2009fast}{{19}{2009}{{Sutton et~al.}}{{Sutton, Maei, Precup, Bhatnagar, Silver, Szepesv{\'a}ri, and Wiewiora}}} \bibcite{Sutton2018book}{{19}{2018}{{Sutton and Barto}}{{}}}
\bibcite{sutton1988learning}{{20}{1988}{{Sutton}}{{}}} \bibcite{sutton2008convergent}{{20}{2008}{{Sutton, Maei, and Szepesv{\'a}ri}}{{}}}
\bibcite{Sutton2018book}{{21}{2018}{{Sutton and Barto}}{{}}} \bibcite{sutton2016emphatic}{{21}{2016}{{Sutton, Mahmood, and White}}{{}}}
\bibcite{sutton2008convergent}{{22}{2008}{{Sutton, Maei, and Szepesv{\'a}ri}}{{}}} \bibcite{tsitsiklis1997analysis}{{22}{1997}{{Tsitsiklis and Van~Roy}}{{}}}
\bibcite{sutton2016emphatic}{{23}{2016}{{Sutton, Mahmood, and White}}{{}}} \bibcite{xu2019reanalysis}{{23}{2019}{{Xu et~al.}}{{Xu, Wang, Zhou, and Liang}}}
\bibcite{tsitsiklis1997analysis}{{24}{1997}{{Tsitsiklis and Van~Roy}}{{}}} \bibcite{zhang2022truncated}{{24}{2022}{{Zhang and Whiteson}}{{}}}
\bibcite{xu2019reanalysis}{{25}{2019}{{Xu et~al.}}{{Xu, Wang, Zhou, and Liang}}}
\bibcite{zhang2022truncated}{{26}{2022}{{Zhang and Whiteson}}{{}}}
\gdef \@abspage@last{8} \gdef \@abspage@last{8}
NEW_aaai/anonymous-submission-latex-2025.bbl
\begin{thebibliography}{26} \begin{thebibliography}{24}
\providecommand{\natexlab}[1]{#1} \providecommand{\natexlab}[1]{#1}
\bibitem[{Baird et~al.(1995)}]{baird1995residual} \bibitem[{Baird et~al.(1995)}]{baird1995residual}
...@@ -16,11 +16,6 @@ Borkar, V.~S. 1997. ...@@ -16,11 +16,6 @@ Borkar, V.~S. 1997.
\newblock Stochastic approximation with two time scales. \newblock Stochastic approximation with two time scales.
\newblock \emph{Syst. \& Control Letters}, 29(5): 291--294. \newblock \emph{Syst. \& Control Letters}, 29(5): 291--294.
\bibitem[{Borkar and Meyn(2000)}]{borkar2000ode}
Borkar, V.~S.; and Meyn, S.~P. 2000.
\newblock The ODE method for convergence of stochastic approximation and reinforcement learning.
\newblock \emph{SIAM J. Control Optim.}, 38(2): 447--469.
\bibitem[{Chen et~al.(2023)Chen, Ma, Li, Yang, Yang, and Gao}]{chen2023modified} \bibitem[{Chen et~al.(2023)Chen, Ma, Li, Yang, Yang, and Gao}]{chen2023modified}
Chen, X.; Ma, X.; Li, Y.; Yang, G.; Yang, S.; and Gao, Y. 2023. Chen, X.; Ma, X.; Li, Y.; Yang, G.; Yang, S.; and Gao, Y. 2023.
\newblock Modified Retrace for Off-Policy Temporal Difference Learning. \newblock Modified Retrace for Off-Policy Temporal Difference Learning.
...@@ -51,11 +46,6 @@ Hallak, A.; Tamar, A.; Munos, R.; and Mannor, S. 2016. ...@@ -51,11 +46,6 @@ Hallak, A.; Tamar, A.; Munos, R.; and Mannor, S. 2016.
\newblock Generalized emphatic temporal difference learning: bias-variance analysis. \newblock Generalized emphatic temporal difference learning: bias-variance analysis.
\newblock In \emph{Proceedings of the 30th AAAI Conference on Artificial Intelligence}, 1631--1637. \newblock In \emph{Proceedings of the 30th AAAI Conference on Artificial Intelligence}, 1631--1637.
\bibitem[{Hirsch(1989)}]{hirsch1989convergent}
Hirsch, M.~W. 1989.
\newblock Convergent activation dynamics in continuous time networks.
\newblock \emph{Neural Netw.}, 2(5): 331--349.
\bibitem[{Johnson and Zhang(2013)}]{johnson2013accelerating} \bibitem[{Johnson and Zhang(2013)}]{johnson2013accelerating}
Johnson, R.; and Zhang, T. 2013. Johnson, R.; and Zhang, T. 2013.
\newblock Accelerating stochastic gradient descent using predictive variance reduction. \newblock Accelerating stochastic gradient descent using predictive variance reduction.
......
NEW_aaai/anonymous-submission-latex-2025.blg
...@@ -3,44 +3,44 @@ Capacity: max_strings=200000, hash_size=200000, hash_prime=170003 ...@@ -3,44 +3,44 @@ Capacity: max_strings=200000, hash_size=200000, hash_prime=170003
The top-level auxiliary file: anonymous-submission-latex-2025.aux The top-level auxiliary file: anonymous-submission-latex-2025.aux
The style file: aaai25.bst The style file: aaai25.bst
Database file #1: aaai25.bib Database file #1: aaai25.bib
You've used 26 entries, You've used 24 entries,
2840 wiz_defined-function locations, 2840 wiz_defined-function locations,
737 strings with 9168 characters, 723 strings with 8880 characters,
and the built_in function-call counts, 19179 in all, are: and the built_in function-call counts, 18055 in all, are:
= -- 1644 = -- 1547
> -- 870 > -- 832
< -- 0 < -- 0
+ -- 321 + -- 305
- -- 288 - -- 276
* -- 1273 * -- 1196
:= -- 2961 := -- 2777
add.period$ -- 107 add.period$ -- 99
call.type$ -- 26 call.type$ -- 24
change.case$ -- 217 change.case$ -- 206
chr.to.int$ -- 27 chr.to.int$ -- 25
cite$ -- 26 cite$ -- 24
duplicate$ -- 1316 duplicate$ -- 1237
empty$ -- 1372 empty$ -- 1285
format.name$ -- 353 format.name$ -- 338
if$ -- 3900 if$ -- 3685
int.to.chr$ -- 1 int.to.chr$ -- 1
int.to.str$ -- 1 int.to.str$ -- 1
missing$ -- 261 missing$ -- 244
newline$ -- 134 newline$ -- 124
num.names$ -- 104 num.names$ -- 96
pop$ -- 614 pop$ -- 586
preamble$ -- 1 preamble$ -- 1
purify$ -- 182 purify$ -- 171
quote$ -- 0 quote$ -- 0
skip$ -- 694 skip$ -- 664
stack$ -- 0 stack$ -- 0
substring$ -- 1043 substring$ -- 969
swap$ -- 703 swap$ -- 658
text.length$ -- 0 text.length$ -- 0
text.prefix$ -- 0 text.prefix$ -- 0
top$ -- 0 top$ -- 0
type$ -- 231 type$ -- 213
warning$ -- 0 warning$ -- 0
while$ -- 166 while$ -- 154
width$ -- 0 width$ -- 0
write$ -- 343 write$ -- 317
NEW_aaai/anonymous-submission-latex-2025.log
This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.3.31) 12 AUG 2024 17:12 This is pdfTeX, Version 3.141592653-2.6-1.40.25 (TeX Live 2023) (preloaded format=pdflatex 2023.3.31) 14 AUG 2024 06:11
entering extended mode entering extended mode
restricted \write18 enabled. restricted \write18 enabled.
file:line:error style messages enabled. file:line:error style messages enabled.
...@@ -581,15 +581,15 @@ Package newfloat Info: `float' package detected. ...@@ -581,15 +581,15 @@ Package newfloat Info: `float' package detected.
\c@eqfn=\count330 \c@eqfn=\count330
\titlearea=\box76 \titlearea=\box76
\actualheight=\skip75 \actualheight=\skip75
LaTeX Font Info: Trying to load font information for U+msa on input line 198. LaTeX Font Info: Trying to load font information for U+msa on input line 213.
(d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsa.fd (d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsa.fd
File: umsa.fd 2013/01/14 v3.01 AMS symbols A File: umsa.fd 2013/01/14 v3.01 AMS symbols A
) )
LaTeX Font Info: Trying to load font information for U+msb on input line 198. LaTeX Font Info: Trying to load font information for U+msb on input line 213.
(d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsb.fd (d:/software/texlive/2023/texmf-dist/tex/latex/amsfonts/umsb.fd
File: umsb.fd 2013/01/14 v3.01 AMS symbols B File: umsb.fd 2013/01/14 v3.01 AMS symbols B
) )
LaTeX Font Info: Trying to load font information for U+esvect on input line 198. LaTeX Font Info: Trying to load font information for U+esvect on input line 213.
(d:/software/texlive/2023/texmf-dist/tex/latex/esvect/uesvect.fd (d:/software/texlive/2023/texmf-dist/tex/latex/esvect/uesvect.fd
File: uesvect.fd File: uesvect.fd
) (./main/introduction.tex ) (./main/introduction.tex
...@@ -597,89 +597,120 @@ Underfull \hbox (badness 3884) in paragraph at lines 39--52 ...@@ -597,89 +597,120 @@ Underfull \hbox (badness 3884) in paragraph at lines 39--52
[]\OT1/ptm/m/n/10 Algorithm sta-bil-ity is promi-nently re-flected in the []\OT1/ptm/m/n/10 Algorithm sta-bil-ity is promi-nently re-flected in the
[] []
LaTeX Font Info: Trying to load font information for TS1+ptm on input line 90.
(d:/software/texlive/2023/texmf-dist/tex/latex/psnfss/ts1ptm.fd
File: ts1ptm.fd 2001/06/04 font definitions for TS1/ptm.
)
Underfull \hbox (badness 1320) in paragraph at lines 90--91
[]\OT1/ptm/m/n/10 Introduction of novel ob-jec-tive func-tions, VBE and
[]
[1{d:/software/texlive/2023/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{d:/software/texlive/2023/texmf-dist/fonts/enc/dvips/base/8r.enc} [1{d:/software/texlive/2023/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{d:/software/texlive/2023/texmf-dist/fonts/enc/dvips/base/8r.enc}
]) (./main/preliminaries.tex [2]) (./main/motivation.tex [3] ]) (./main/preliminaries.tex
Overfull \hbox (19.88512pt too wide) detected at line 90
[]
[]
LaTeX Warning: Command \textellipsis invalid in math mode on input line 23.
Underfull \hbox (badness 4120) in paragraph at lines 140--149 LaTeX Font Info: Trying to load font information for OMS+ptm on input line 58.
\OT1/ptm/m/n/10 a given set of sub-samples|in the form of triples (d:/software/texlive/2023/texmf-dist/tex/latex/psnfss/omsptm.fd
File: omsptm.fd
)
LaTeX Font Info: Font shape `OMS/ptm/m/sc' in size <9> not available
(Font) Font shape `OMS/cmsy/m/n' tried instead on input line 58.
<main/pic/2StateExample.pdf, id=13, 448.67625pt x 158.09062pt>
File: main/pic/2StateExample.pdf Graphic file (type pdf)
<use main/pic/2StateExample.pdf>
Package pdftex.def Info: main/pic/2StateExample.pdf used on input line 98.
(pdftex.def) Requested size: 71.81705pt x 35.90657pt.
[2 <./main/pic/2StateExample.pdf>]
Underfull \hbox (badness 1472) in paragraph at lines 255--262
\OT1/ptm/m/n/10 where $[]$ is a di-ag-o-nal ma-trix with di-ag-o-nal el-e-ments
[] []
Overfull \hbox (32.7115pt too wide) detected at line 184 Underfull \hbox (badness 1931) in paragraph at lines 282--290
[] []\OT1/ptm/m/n/10 Minimum eigen-value larger, al-go-rithm's con-ver-gence
[] []
) (./main/motivation.tex
Overfull \hbox (38.64899pt too wide) detected at line 264 Overfull \hbox (36.0123pt too wide) detected at line 119
[] []
[] []
[3]
Overfull \hbox (30.29318pt too wide) detected at line 287 Overfull \hbox (13.08682pt too wide) detected at line 182
[] []
[] []
[4]
Overfull \hbox (19.45059pt too wide) detected at line 300
[]
[]
) (./main/theory.tex [5] Overfull \hbox (50.91507pt too wide) detected at line 381
Underfull \hbox (badness 2035) in paragraph at lines 140--145 []
\OT1/ptm/m/n/10 2000) are ver-i-fied. Fur-ther-more, As-sump-tions (TS) of
[] []
<main/pic/2-state.pdf, id=41, 586.83241pt x 443.7378pt> ) (./main/theory.tex [4] [5]) (./main/experiment.tex
File: main/pic/2-state.pdf Graphic file (type pdf) <main/pic/maze_13_13.pdf, id=51, 493.1646pt x 387.62602pt>
<use main/pic/2-state.pdf> File: main/pic/maze_13_13.pdf Graphic file (type pdf)
Package pdftex.def Info: main/pic/2-state.pdf used on input line 483. <use main/pic/maze_13_13.pdf>
(pdftex.def) Requested size: 131.66446pt x 113.71024pt. Package pdftex.def Info: main/pic/maze_13_13.pdf used on input line 4.
<main/pic/7-state.pdf, id=42, 593.3367pt x 452.4102pt> (pdftex.def) Requested size: 98.63116pt x 77.52382pt.
File: main/pic/7-state.pdf Graphic file (type pdf) <main/pic/2-state-onpolicy.pdf, id=52, 590.44589pt x 448.07399pt>
<use main/pic/7-state.pdf> File: main/pic/2-state-onpolicy.pdf Graphic file (type pdf)
Package pdftex.def Info: main/pic/7-state.pdf used on input line 487. <use main/pic/2-state-onpolicy.pdf>
(pdftex.def) Requested size: 131.66617pt x 113.70975pt. Package pdftex.def Info: main/pic/2-state-onpolicy.pdf used on input line 13.
<main/pic/maze.pdf, id=43, 578.16pt x 505.89pt> (pdftex.def) Requested size: 155.6025pt x 138.84679pt.
<main/pic/2-state-offpolicy.pdf, id=53, 596.2275pt x 462.52798pt>
File: main/pic/2-state-offpolicy.pdf Graphic file (type pdf)
<use main/pic/2-state-offpolicy.pdf>
Package pdftex.def Info: main/pic/2-state-offpolicy.pdf used on input line 17.
(pdftex.def) Requested size: 155.59775pt x 138.85121pt.
<main/pic/maze.pdf, id=54, 578.16pt x 505.89pt>
File: main/pic/maze.pdf Graphic file (type pdf) File: main/pic/maze.pdf Graphic file (type pdf)
<use main/pic/maze.pdf> <use main/pic/maze.pdf>
Package pdftex.def Info: main/pic/maze.pdf used on input line 491. Package pdftex.def Info: main/pic/maze.pdf used on input line 21.
(pdftex.def) Requested size: 131.65952pt x 113.71227pt. (pdftex.def) Requested size: 155.60243pt x 138.84613pt.
<main/pic/cl.pdf, id=44, 578.16pt x 505.89pt> <main/pic/cl.pdf, id=55, 578.16pt x 505.89pt>
File: main/pic/cl.pdf Graphic file (type pdf) File: main/pic/cl.pdf Graphic file (type pdf)
<use main/pic/cl.pdf> <use main/pic/cl.pdf>
Package pdftex.def Info: main/pic/cl.pdf used on input line 495. Package pdftex.def Info: main/pic/cl.pdf used on input line 25.
(pdftex.def) Requested size: 131.65952pt x 113.71227pt. (pdftex.def) Requested size: 155.60243pt x 138.84613pt.
<main/pic/mt.pdf, id=45, 578.16pt x 505.89pt> <main/pic/mt.pdf, id=56, 578.16pt x 505.89pt>
File: main/pic/mt.pdf Graphic file (type pdf) File: main/pic/mt.pdf Graphic file (type pdf)
<use main/pic/mt.pdf> <use main/pic/mt.pdf>
Package pdftex.def Info: main/pic/mt.pdf used on input line 499. Package pdftex.def Info: main/pic/mt.pdf used on input line 29.
(pdftex.def) Requested size: 131.65952pt x 113.71227pt. (pdftex.def) Requested size: 155.60243pt x 138.84613pt.
<main/pic/acrobot.pdf, id=46, 578.16pt x 505.89pt> <main/pic/acrobot.pdf, id=57, 578.16pt x 505.89pt>
File: main/pic/acrobot.pdf Graphic file (type pdf) File: main/pic/acrobot.pdf Graphic file (type pdf)
<use main/pic/acrobot.pdf> <use main/pic/acrobot.pdf>
Package pdftex.def Info: main/pic/acrobot.pdf used on input line 503. Package pdftex.def Info: main/pic/acrobot.pdf used on input line 33.
(pdftex.def) Requested size: 131.65952pt x 113.71227pt. (pdftex.def) Requested size: 155.60243pt x 138.84613pt.
[6]) (./main/experiment.tex) (./main/conclusion.tex) (./anonymous-submission-latex-2025.bbl [7 <./main/pic/2-state.pdf> <./main/pic/7-state.pdf> <./main/pic/maze.pdf> <./main/pic/cl.pdf> <./main/pic/mt.pdf> <./main/pic/acrobot.pdf>]) [8] (./anonymous-submission-latex-2025.aux) )
LaTeX Warning: `h' float specifier changed to `ht'.
[6 <./main/pic/maze_13_13.pdf>]
Underfull \vbox (badness 10000) has occurred while \output is active []
) (./main/conclusion.tex
Underfull \hbox (badness 2951) in paragraph at lines 32--33
[]\OT1/ptm/m/n/10 analysis of the con-ver-gence rate of VMTDC and
[]
) (./anonymous-submission-latex-2025.bbl [7 <./main/pic/2-state-onpolicy.pdf> <./main/pic/2-state-offpolicy.pdf> <./main/pic/maze.pdf> <./main/pic/cl.pdf> <./main/pic/mt.pdf> <./main/pic/acrobot.pdf>]) [8] (./anonymous-submission-latex-2025.aux) )
Here is how much of TeX's memory you used: Here is how much of TeX's memory you used:
18678 strings out of 476025 18693 strings out of 476025
360333 string characters out of 5789524 361060 string characters out of 5789524
1897382 words of memory out of 5000000 1890382 words of memory out of 5000000
38742 multiletter control sequences out of 15000+600000 38746 multiletter control sequences out of 15000+600000
546921 words of font info for 105 fonts, out of 8000000 for 9000 549956 words of font info for 109 fonts, out of 8000000 for 9000
1141 hyphenation exceptions out of 8191 1141 hyphenation exceptions out of 8191
84i,17n,89p,423b,663s stack positions out of 10000i,1000n,20000p,200000b,200000s 84i,17n,89p,423b,419s stack positions out of 10000i,1000n,20000p,200000b,200000s
<d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmib10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cmextra/cmmib7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmb8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmbi8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmr8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmri8a.pfb> <d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmmib10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmr9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy5.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy6.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy9.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmb8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmbi8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmr8a.pfb><d:/software/texlive/2023/texmf-dist/fonts/type1/urw/times/utmri8a.pfb>
Output written on anonymous-submission-latex-2025.pdf (8 pages, 1216302 bytes). Output written on anonymous-submission-latex-2025.pdf (8 pages, 1896294 bytes).
PDF statistics: PDF statistics:
345 PDF objects out of 1000 (max. 8388607) 375 PDF objects out of 1000 (max. 8388607)
136 compressed objects within 2 object streams 147 compressed objects within 2 object streams
0 named destinations out of 1000 (max. 500000) 0 named destinations out of 1000 (max. 500000)
43 words of extra memory for PDF output out of 10000 (max. 10000000) 53 words of extra memory for PDF output out of 10000 (max. 10000000)
NEW_aaai/anonymous-submission-latex-2025.tex
...@@ -119,7 +119,7 @@ ...@@ -119,7 +119,7 @@
% nouns, adverbs, adjectives should be capitalized, including both words in hyphenated terms, while % nouns, adverbs, adjectives should be capitalized, including both words in hyphenated terms, while
% articles, conjunctions, and prepositions are lower case unless they % articles, conjunctions, and prepositions are lower case unless they
% directly follow a colon or long dash % directly follow a colon or long dash
\title{A Variance Minimization Approach to Off-policy Temporal-Difference Learning} \title{A Variance Minimization Approach to Temporal-Difference Learning}
\author{ \author{
%Authors %Authors
% All authors must be in the same font size and format. % All authors must be in the same font size and format.
...@@ -194,16 +194,31 @@ ...@@ -194,16 +194,31 @@
\maketitle \maketitle
% \setcounter{theorem}{0} % \setcounter{theorem}{0}
\begin{abstract} \begin{abstract}
In this paper, we introduce the concept of improving the performance of parametric % In this paper, we introduce the concept of improving the performance of parametric
Temporal-Difference (TD) learning algorithms by the Variance Minimization (VM) parameter, $\omega$, % Temporal-Difference (TD) learning algorithms by the Variance Minimization (VM) parameter, $\omega$,
which is dynamically updated at each time step. Specifically, we incorporate the VM parameter into off-policy linear algorithms such as TDC and ETD, resulting in the % which is dynamically updated at each time step. Specifically, we incorporate the VM parameter into off-policy linear algorithms such as TDC and ETD, resulting in the
Variance Minimization TDC (VMTDC) algorithm and the Variance Minimization ETD (VMETD) algorithm. In the two-state counterexample, % Variance Minimization TDC (VMTDC) algorithm and the Variance Minimization ETD (VMETD) algorithm. In the two-state counterexample,
% we analyze
% the convergence speed of these algorithms by calculating the minimum eigenvalue of the key
% matrices and find that the VMTDC algorithm converges faster than TDC, while VMETD is more stable in convergence than ETD
% through the
% experiment.In controlled experiments, the VM algorithms demonstrate
% superior performance.
Under certain conditions, the larger the smallest
eigenvalue of the key matrix of an algorithm, the
faster the algorithm converges. By observation, most
current objective functions aim to minimize error.
Therefore, in this paper, we propose two new objective
functions and derive three Variance Minimization (VM) algorithms, including VMTD, VMTDC and VMETD.
A scalar parameter, $\omega$, is introduced, to improve the performance of parametric
Temporal-Difference (TD) learning algorithms.
In the policy evaluation experiment, two-state,
we analyze we analyze
the convergence speed of these algorithms by calculating the minimum eigenvalue of the key the convergence speed of these algorithms by calculating the minimum eigenvalue of the key
matrices and find that the VMTDC algorithm converges faster than TDC, while VMETD is more stable in convergence than ETD matrices both on-policy and off-policy.In controlled experiments, the VM algorithms demonstrate
through the
experiment.In controlled experiments, the VM algorithms demonstrate
superior performance. superior performance.
\end{abstract} \end{abstract}
% Uncomment the following to link to your code, datasets, an extended version or similar. % Uncomment the following to link to your code, datasets, an extended version or similar.
......
NEW_aaai/main/conclusion.tex
\section{Conclusion and Future Work} \section{Conclusion and Future Work}
% Value-based reinforcement learning typically aims
% to minimize error as an optimization objective.
% As an alternation, this study proposes new objective
% functions: VBE and VPBE, and derives many variance minimization algorithms, including VMTD,
% VMTDC and VMETD.
% All algorithms demonstrated superior performance in policy
% evaluation and control experiments.
% Future work may include, but are not limited
% to, (1) analysis of the convergence rate of VMTDC and VMETD.
% (2) extensions of VBE and VPBE to multi-step returns.
% (3) extensions to nonlinear approximations, such as neural networks.
Value-based reinforcement learning typically aims Value-based reinforcement learning typically aims
to minimize error as an optimization objective. to minimize error as an optimization objective.
As an alternation, this study proposes new objective As an alternation, this study proposes two new objective
functions: VBE and VPBE, and derives many variance minimization algorithms, including VMTD, functions: VBE and VPBE, and derives an on-policy algorithm:
VMTDC and VMETD. VMTD and two off-policy algorithms: VMTDC and VMETD.
% The VMTD algorithm
% is essentially an adjustment or correction to the traditional
% TD update.
% Both
% algorithms are capable of stabilizing gradient estimation, reducing
% the variance of gradient estimation and accelerating convergence.
All algorithms demonstrated superior performance in policy All algorithms demonstrated superior performance in policy
evaluation and control experiments. evaluation and control experiments.
Both algorithms demonstrated superior performance in policy
evaluation and control experiments.
Future work may include, but are not limited Future work may include, but are not limited
to, (1) analysis of the convergence rate of VMTDC and VMETD. to,
(2) extensions of VBE and VPBE to multi-step returns. \begin{itemize}
(3) extensions to nonlinear approximations, such as neural networks. \item analysis of the convergence rate of VMTDC and VMETD.
\ No newline at end of file \item extensions of VBE and VPBE to multi-step returns.
\item extensions to nonlinear approximations, such as neural networks.
\end{itemize}
\ No newline at end of file
NEW_aaai/main/experiment.tex
% \subsection{Testing Tasks}
\begin{figure}[h]
\centering
\includegraphics[scale=0.2]{main/pic/maze_13_13.pdf}
\caption{Maze.}
\end{figure}
\begin{figure*}[tb]
\vskip 0.2in
\begin{center}
\subfigure[on-policy 2-state]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/2-state-onpolicy.pdf}
\label{2-state}
}
\subfigure[off-policy 2-state]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/2-state-offpolicy.pdf}
\label{7-state}
}
\subfigure[Maze]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/maze.pdf}
\label{MazeFull}
}\\
\subfigure[Cliff Walking]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/cl.pdf}
\label{CliffWalkingFull}
}
\subfigure[Mountain Car]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/mt.pdf}
\label{MountainCarFull}
}
\subfigure[Acrobot]{
\includegraphics[width=0.65\columnwidth, height=0.58\columnwidth]{main/pic/acrobot.pdf}
\label{AcrobotFull}
}
\caption{Learning curses of one evaluation environment and four contral environments.}
\label{Complete_full}
\end{center}
\vskip -0.2in
\end{figure*}
\section{Experimental Studies} \section{Experimental Studies}
This section assesses algorithm performance through experiments, This section assesses algorithm performance through experiments,
which are divided into policy evaluation experiments and control experiments. which are divided into policy evaluation experiments and control experiments.
The control algorithms for TDC, ETD, VMTDC, and VMETD are named GQ, EQ, VMGQ, and VMEQ, respectively. The evaluation experimental environments is the 2-state.
The evaluation experimental environments are the 2-state and 7-state counterexample. In a 2-state environment, we conducted two types of experiments—on-policy
and off-policy—to verify the relationship between the convergence speed of
the algorithm and the smallest eigenvalue of the key matrix $\textbf{A}$.
Control experiments, by allowing the algorithm to interact
with the environment to optimize the policy, can evaluate its
performance in learning the optimal policy. This provides a more
comprehensive assessment of the algorithm's overall capabilities.
The control experimental environments are Maze, CliffWalking-v0, MountainCar-v0, and Acrobot-v1. The control experimental environments are Maze, CliffWalking-v0, MountainCar-v0, and Acrobot-v1.
The control algorithms for TDC, ETD, VMTDC, and VMETD are named GQ, EQ, VMGQ, and VMEQ, respectively.
For TD and VMTD control algorithms, there are two variants each: Sarsa and Q-learning for TD, and VMSarsa and VMQ for VMTD.
% For specific experimental parameters, please refer to the appendix.
% \textbf{Baird's off-policy counterexample:} This task is well known as a
% counterexample, in which TD diverges \cite{baird1995residual,sutton2009fast}. As
% shown in Figure \ref{bairdexample}, reward for each transition is zero. Thus the true values are zeros for all states and for any given policy. The behaviour policy
% chooses actions represented by solid lines with a probability of $\frac{1}{7}$
% and actions represented by dotted lines with a probability of $\frac{6}{7}$. The
% target policy is expected to choose the solid line with more probability than $\frac{1}{7}$,
% and it chooses the solid line with probability of $1$ in this paper.
% The discount factor $\gamma =0.99$, and the feature matrix is
% defined in Appendix \ref{experimentaldetails} \cite{baird1995residual,sutton2009fast,maei2011gradient}.
% \begin{figure}
% \begin{center}
% \input{main/pic/BairdExample.tex}
% \caption{7-state.}
% \label{bairdexample}
% \end{center}
% \end{figure}
% The feature matrix of 7-state version of Baird's off-policy counterexample is
% defined as follow:
% \begin{equation*}
% \Phi_{Counter}=\left[
% \begin{array}{cccccccc}
% 1 & 2& 0& 0& 0& 0& 0& 0\\
% 1 & 0& 2& 0& 0& 0& 0& 0\\
% 1 & 0& 0& 2& 0& 0& 0& 0\\
% 1 & 0& 0& 0& 2& 0& 0& 0\\
% 1 & 0& 0& 0& 0& 2& 0& 0\\
% 1 & 0& 0& 0& 0& 0& 2& 0\\
% 2 & 0& 0& 0& 0& 0& 0& 1
% \end{array}\right]
% \end{equation*}
\subsection{Testing Tasks}
% \begin{figure}[h]
% \centering
% \includegraphics[scale=0.2]{main/pic/maze_13_13.pdf}
% \caption{Maze.}
% \end{figure}
\textbf{Maze}: The learning agent should find a shortest path from the upper
left corner to the lower right corner.
In each state,
there are four alternative actions: $up$, $down$, $left$, and $right$, which
takes the agent deterministically to the corresponding neighbour state,
except when a movement is blocked by an obstacle or the edge
of the maze. Rewards are $-1$ in all transitions until the
agent reaches the goal state.
The discount factor $\gamma=0.99$, and states $s$ are represented by tabular
features.The maximum number of moves in the game is set to 1000.
\textbf{The other three control environments}: Cliff Walking, Mountain Car, and Acrobot are
selected from the gym official website and correspond to the following
versions: ``CliffWalking-v0'', ``MountainCar-v0'' and ``Acrobot-v1''.
For specific details, please refer to the gym official website.
The maximum number of steps for the Mountain Car environment is set to 1000,
while the default settings are used for the other two environments. In Mountain car and Acrobot, features are generated by tile coding.
For all policy evaluation experiments, each experiment
is independently run 100 times.
For all control experiments, each experiment is independently run 50 times.
For specific experimental parameters, please refer to the appendix. For specific experimental parameters, please refer to the appendix.
For the evaluation experiment, the experimental results \subsection{Experimental Results and Analysis}
align with our previous analysis. In the 2-state counterexample Figure \ref{2-state} shows the learning curves for the on-policy
environment, the TDC algorithm has the smallest minimum 2-state policy evaluation experiment. In this setup,
eigenvalue of the key matrix, resulting in the slowest the convergence speed of TD, VMTD, TDC, and VMTDC decreases
convergence speed. In contrast, the minimum eigenvalue sequentially. Table \ref{tab:min_eigenvalues} indicates that the smallest eigenvalue
of VMTDC is larger, leading to faster convergence. of the key matrix for these four algorithms is greater than 0
Although VMETD's minimum eigenvalue is larger than ETD's, and decreases sequentially, which is consistent with the
causing VMETD to converge more slowly than ETD in the experimental curves and table values.
2-state counterexample, the standard deviation (shaded area)
of VMETD is smaller than that of ETD, indicating that VMETD Figure B displays the learning curves for the off-policy
converges more smoothly. In the 7-state counterexample 2-state policy evaluation experiment. In this setup,
environment, VMTDC converges faster than TDC and both VMETD and ETD are diverge. the convergence speed of ETD, VMETD, VMTD, VMTDC, and
TDC decreases sequentially, while TD diverges. Table \ref{tab:min_eigenvalues}
For the control experiments, the results for the maze and shows that the smallest eigenvalue of the key matrix for
cliff walking environments are similar: VMGQ ETD, VMETD, VMTD, VMTDC, and TDC is greater than 0 and
outperforms GQ, EQ outperforms VMGQ, and VMEQ performs decreases sequentially, while the smallest eigenvalue
the best. In the mountain car and Acrobot experiments, for TD is less than 0. This is consistent with the
VMGQ and VMEQ show comparable performance, both outperforming experimental curves and table values. Remarkably,
GQ and EQ. In summary, for control experiments, VM algorithms although VMTD is guaranteed to converge under
outperform non-VM algorithms. on-policy conditions, it still converges in the
off-policy 2-state scenario. The update formula
In summary, the performance of VMSarsa, of VMTD indicates that it is essentially an
VMQ, and VMGQ(0) is better than that of other algorithms. adjustment and correction of the TD update,
In the Cliff Walking environment, with the introduction of the parameter $\omega$
the performance of VMGQ(0) is slightly better than that of making the variance of the gradient estimate
VMSarsa and VMQ. In the other three experimental environments, more stable, thereby making the update of theta more stable.
the performances of VMSarsa, VMQ, and VMGQ(0) are close.
\ No newline at end of file Figures \ref{MazeFull}, \ref{CliffWalkingFull}, \ref{MountainCarFull} and \ref{AcrobotFull} show the learning curves
for four control experiments. A common feature
observed across these experiments is that VMEQ
outperforms EQ, VMGQ outperforms GQ, VMQ outperforms
Q-learning, and VMSarsa outperforms Sarsa. For the
Maze and Cliffwalking experiments, VMEQ demonstrated
the best performance with the fastest convergence speed.
In the Mountain Car and Acrobot experiments, the performance
of the four VM algorithms was nearly identical and all
outperformed the other algorithms.
Overall, whether in policy evaluation experiments or
control experiments, the VM algorithms have
demonstrated superior performance,
especially excelling in the control experiments.
\ No newline at end of file
NEW_aaai/main/introduction.tex
...@@ -68,26 +68,28 @@ based on recursive optimization using it are known to be unstable. ...@@ -68,26 +68,28 @@ based on recursive optimization using it are known to be unstable.
It is necessary to propose a new objective function, but the mentioned objective functions above are all some form of error. It is necessary to propose a new objective function, but the mentioned objective functions above are all some form of error.
Is minimizing error the only option for value-based reinforcement learning? Is minimizing error the only option for value-based reinforcement learning?
For policy evaluation experiments, % For policy evaluation experiments,
differences in objective functions may result % differences in objective functions may result
in inconsistent fixed points. This inconsistency % in inconsistent fixed points. This inconsistency
makes it difficult to uniformly compare the superiority % makes it difficult to uniformly compare the superiority
of algorithms derived from different objective functions. % of algorithms derived from different objective functions.
However, for control experiments, since the choice of actions % However, for control experiments, since the choice of actions
depends on the relative values of the Q values rather than their % depends on the relative values of the Q values rather than their
absolute values, the presence of solution bias is acceptable. % absolute values, the presence of solution bias is acceptable.
Based on this observation, we propose alternate objective functions Based on this observation, we propose alternate objective functions
instead of minimizing errors. We minimize instead of minimizing errors. We minimize Variance of Bellman Error (VBE) and
Variance of Projected Bellman Error (VPBE) Variance of Projected Bellman Error (VPBE)
and derive Variance Minimization (VM) algorithms. and derive Variance Minimization (VM) algorithms.
These algorithms preserve the invariance of the optimal policy in the control environments, These algorithms preserve the invariance of the optimal policy in the control environments,
but significantly reduce the variance of gradient estimation, and significantly reduce the variance of gradient estimation,
and thus hastening convergence. and thus hastening convergence.
The contributions of this paper are as follows: The contributions of this paper are as follows:
(1) Introduction of novel objective functions based on \begin{itemize}
the invariance of the optimal policy. \item Introduction of novel objective functions, VBE and VPBE.
(2) Propose two off-policy variance minimization algorithms. \item Propose a on-policy VM algorithm and two off-policy VM algorithms.
(3) Proof of their convergence. \item Proof of their convergence.
(5) Experiments demonstrating the faster convergence speed of the proposed algorithms. \item The experiments demonstrate the superiority of the VM algorithms.
\end{itemize}
NEW_aaai/main/motivation.tex
\section{Variance Minimization Algorithms} \section{Variance Minimization Algorithms}
To derive an algorithm with a larger minimum eigenvalue for matrix This section will introduce two new objective functions and
$\textbf{A}$, it is necessary to propose new objective functions. three new algorithms, including one on-policy algorithm and two off-policy algorithms, and calculate the minimum eigenvalue
The mentioned objective functions in the Introduction of $\textbf{A}$ for each of the three algorithms under on-policy and
are all forms of error. Is minimizing error the only option off-policy in a 2-state environment, thereby comparing the
for value-based reinforcement learning? Based on this observation, convergence speed of the three algorithms.
we propose alternative objective functions instead of minimizing errors. % To derive an algorithm with a larger minimum eigenvalue for matrix
We minimize the Variance of Projected Bellman Error (VPBE) and derive the % $\textbf{A}$, it is necessary to propose new objective functions.
VMTDC algorithm. This idea is then innovatively applied to ETD, resulting % The mentioned objective functions in the Introduction
in the VMETD algorithm. % are all forms of error. Is minimizing error the only option
% for value-based reinforcement learning? Based on this observation,
% We propose alternative objective functions instead of minimizing errors.
% We minimize the Variance of Projected Bellman Error (VPBE) and derive the
% VMTDC algorithm. This idea is then innovatively applied to ETD, resulting
% in the VMETD algorithm.
% \subsection{Motivation} % \subsection{Motivation}
% gagagga % gagagga
\begin{algorithm}[t] % \begin{algorithm}[t]
\caption{VMTDC algorithm with linear function approximation in the off-policy setting} % \caption{VMTDC algorithm with linear function approximation in the off-policy setting}
\label{alg:algorithm 2} % \label{alg:algorithm 2}
\begin{algorithmic} % \begin{algorithmic}
\STATE {\bfseries Input:} $\bm{\theta}_{0}$, $\bm{u}_0$, $\omega_{0}$, $\gamma % \STATE {\bfseries Input:} $\bm{\theta}_{0}$, $\bm{u}_0$, $\omega_{0}$, $\gamma
$, learning rate $\alpha_t$, $\zeta_t$ and $\beta_t$, behavior policy $\mu$ and target policy $\pi$ % $, learning rate $\alpha_t$, $\zeta_t$ and $\beta_t$, behavior policy $\mu$ and target policy $\pi$
\REPEAT % \REPEAT
\STATE For any episode, initialize $\bm{\theta}_{0}$ arbitrarily, $\bm{u}_{0}$ and $\omega_{0}$ to $0$, $\gamma \in (0,1]$, and $\alpha_t$, $\zeta_t$ and $\beta_t$ are constant.\\ % \STATE For any episode, initialize $\bm{\theta}_{0}$ arbitrarily, $\bm{u}_{0}$ and $\omega_{0}$ to $0$, $\gamma \in (0,1]$, and $\alpha_t$, $\zeta_t$ and $\beta_t$ are constant.\\
% \textbf{Output}: $\bm{\theta}^*$.\\ % % \textbf{Output}: $\bm{\theta}^*$.\\
\FOR{$t=0$ {\bfseries to} $T-1$} % \FOR{$t=0$ {\bfseries to} $T-1$}
\STATE Take $A_t$ from $S_t$ according to $\mu$, and arrive at $S_{t+1}$\\ % \STATE Take $A_t$ from $S_t$ according to $\mu$, and arrive at $S_{t+1}$\\
\STATE Observe sample ($S_t$,$R_{t+1}$,$S_{t+1}$) at time step $t$ (with their corresponding state feature vectors)\\ % \STATE Observe sample ($S_t$,$R_{t+1}$,$S_{t+1}$) at time step $t$ (with their corresponding state feature vectors)\\
\STATE $\delta_t = R_{t+1}+\gamma\bm{\theta}_t^{\top}\bm{\phi}_{t+1}-\bm{\theta}_t^{\top}\bm{\phi}_t$ % \STATE $\delta_t = R_{t+1}+\gamma\bm{\theta}_t^{\top}\bm{\phi}_{t+1}-\bm{\theta}_t^{\top}\bm{\phi}_t$
\STATE $\rho_{t} \leftarrow \frac{\pi(A_t | S_t)}{\mu(A_t | S_t)}$ % \STATE $\rho_{t} \leftarrow \frac{\pi(A_t | S_t)}{\mu(A_t | S_t)}$
\STATE $\bm{\theta}_{t+1}\leftarrow \bm{\theta}_{t}+\alpha_t [ (\rho_t\delta_t-\omega_t)\bm{\phi}_t - \gamma \rho_t\bm{\phi}_{t+1}(\bm{\phi}^{\top}_{t} \bm{u}_{t})]$ % \STATE $\bm{\theta}_{t+1}\leftarrow \bm{\theta}_{t}+\alpha_t [ (\rho_t\delta_t-\omega_t)\bm{\phi}_t - \gamma \rho_t\bm{\phi}_{t+1}(\bm{\phi}^{\top}_{t} \bm{u}_{t})]$
\STATE $\bm{u}_{t+1}\leftarrow \bm{u}_{t}+\zeta_t[(\rho_t\delta_t-\omega_t) - \bm{\phi}^{\top}_{t} \bm{u}_{t}] \bm{\phi}_t$ % \STATE $\bm{u}_{t+1}\leftarrow \bm{u}_{t}+\zeta_t[(\rho_t\delta_t-\omega_t) - \bm{\phi}^{\top}_{t} \bm{u}_{t}] \bm{\phi}_t$
\STATE $\omega_{t+1}\leftarrow \omega_{t}+\beta_t (\rho_t\delta_t-\omega_t)$ % \STATE $\omega_{t+1}\leftarrow \omega_{t}+\beta_t (\rho_t\delta_t-\omega_t)$
\STATE $S_t=S_{t+1}$ % \STATE $S_t=S_{t+1}$
\ENDFOR % \ENDFOR
\UNTIL{terminal episode} % \UNTIL{terminal episode}
\end{algorithmic} % \end{algorithmic}
\end{algorithm} % \end{algorithm}
\begin{algorithm}[t] % \begin{algorithm}[t]
\caption{VMETD algorithm with linear function approximation in the off-policy setting} % \caption{VMETD algorithm with linear function approximation in the off-policy setting}
\label{alg:algorithm 5} % \label{alg:algorithm 5}
\begin{algorithmic} % \begin{algorithmic}
\STATE {\bfseries Input:} $\bm{\theta}_{0}$, $F_0$, $\omega_{0}$, $\gamma % \STATE {\bfseries Input:} $\bm{\theta}_{0}$, $F_0$, $\omega_{0}$, $\gamma
$, learning rate $\alpha_t$, $\zeta_t$ and $\beta_t$, behavior policy $\mu$ and target policy $\pi$ % $, learning rate $\alpha_t$, $\zeta_t$ and $\beta_t$, behavior policy $\mu$ and target policy $\pi$
\REPEAT % \REPEAT
\STATE For any episode, initialize $\bm{\theta}_{0}$ arbitrarily, $F_{0}$ to $1$ and $\omega_{0}$ to $0$, $\gamma \in (0,1]$, and $\alpha_t$, $\zeta_t$ and $\beta_t$ are constant.\\ % \STATE For any episode, initialize $\bm{\theta}_{0}$ arbitrarily, $F_{0}$ to $1$ and $\omega_{0}$ to $0$, $\gamma \in (0,1]$, and $\alpha_t$, $\zeta_t$ and $\beta_t$ are constant.\\
% \textbf{Output}: $\theta^*$.\\ % % \textbf{Output}: $\theta^*$.\\
\FOR{$t=0$ {\bfseries to} $T-1$} % \FOR{$t=0$ {\bfseries to} $T-1$}
\STATE Take $A_t$ from $S_t$ according to $\mu$, and arrive at $S_{t+1}$\\ % \STATE Take $A_t$ from $S_t$ according to $\mu$, and arrive at $S_{t+1}$\\
\STATE Observe sample ($S_t$,$R_{t+1}$,$S_{t+1}$) at time step $t$ (with their corresponding state feature vectors)\\ % \STATE Observe sample ($S_t$,$R_{t+1}$,$S_{t+1}$) at time step $t$ (with their corresponding state feature vectors)\\
\STATE $\delta_t = R_{t+1}+\gamma\bm{\theta}_t^{\top}\bm{\phi}_{t+1}-\bm{\theta}_t^{\top}\bm{\phi}_t$ % \STATE $\delta_t = R_{t+1}+\gamma\bm{\theta}_t^{\top}\bm{\phi}_{t+1}-\bm{\theta}_t^{\top}\bm{\phi}_t$
\STATE $\rho_{t} \leftarrow \frac{\pi(A_t | S_t)}{\mu(A_t | S_t)}$ % \STATE $\rho_{t} \leftarrow \frac{\pi(A_t | S_t)}{\mu(A_t | S_t)}$
\STATE $F_{t}\leftarrow \gamma \rho_t F_{t-1} +1$ % \STATE $F_{t}\leftarrow \gamma \rho_t F_{t-1} +1$
\STATE $\bm{\theta}_{t+1}\leftarrow \bm{\theta}_{t}+\alpha_t (F_t \rho_t\delta_t-\omega_t)\bm{\phi}_t$ % \STATE $\bm{\theta}_{t+1}\leftarrow \bm{\theta}_{t}+\alpha_t (F_t \rho_t\delta_t-\omega_t)\bm{\phi}_t$
\STATE $\omega_{t+1}\leftarrow \omega_{t}+\beta_t (F_t \rho_t\delta_t-\omega_t)$ % \STATE $\omega_{t+1}\leftarrow \omega_{t}+\beta_t (F_t \rho_t\delta_t-\omega_t)$
\STATE $S_t=S_{t+1}$ % \STATE $S_t=S_{t+1}$
\ENDFOR % \ENDFOR
\UNTIL{terminal episode} % \UNTIL{terminal episode}
\end{algorithmic} % \end{algorithmic}
\end{algorithm} % \end{algorithm}
\subsection{Variance Minimization TD Learning: VMTD}
For on-policy learning,
a novel objective function, Variance of Bellman Error (VBE), is proposed as follows:
% \begin{equation}
% \begin{array}{ccl}
% \arg \min_{\theta}\text{VBE}(\theta)&=&\arg \min_{\theta}\mathbb{E}[(\mathbb{E}[\delta|s]-\mathbb{E}[\mathbb{E}[\delta|s]])^2]\\
% &=&\arg \min_{\theta,\omega} \mathbb{E}[(\mathbb{E}[\delta|s]-\omega)^2].
% \end{array}
% \end{equation}
\subsection{Variance Minimization TDC Learning: VMTDC}
For off-policy learning, we propose a new objective function,
called Variance of Projected Bellman error (VPBE),
and the corresponding algorithm is called VMTDC.
\begin{align} \begin{align}
\text{VPBE}(\bm{\theta}) &= \mathbb{E}[(\delta-\mathbb{E}[\delta]) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1} \\ \arg \min_{\theta}\text{VBE}(\theta) &= \arg \min_{\theta}\mathbb{E}[(\mathbb{E}[\delta_t|S_t]-\mathbb{E}[\mathbb{E}[\delta_t|S_t]])^2] \\
& \mathbb{E}[(\delta -\mathbb{E}[\delta ])\bm{\phi}] \notag \\ &= \arg \min_{\theta,\omega} \mathbb{E}[(\mathbb{E}[\delta_t|S_t]-\omega)^2]\notag
&= (\bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}))^{\top}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1} \notag \\
& \bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
&= (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}})^{\top}\textbf{D}^{\top}\bm{\Phi}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1} \notag \\
& \bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
&= (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}})^{\top}\Pi^{\top}\textbf{D}\Pi \notag \\
& (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
&= (\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}-\textbf{W}_{\bm{\theta}}))^{\top}\textbf{D} \notag \\
& (\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}-\textbf{W}_{\bm{\theta}})) \notag \\
&= ||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}} - \textbf{W}_{\bm{\theta}})||^{2}_{\mu} \notag \\
&= ||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}) - \Pi\textbf{W}_{\bm{\theta}}||^{2}_{\mu} \notag \\
&= \mathbb{E}[(\delta-\omega) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\omega)\bm{\phi}]
\end{align} \end{align}
where $\textbf{W}_{\bm{\theta}}$ is viewed as vectors with every element being equal to $||\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}||^{2}_{\mu}$ and $\omega$ is used to approximate $\mathbb{E}[\delta]$, i.e., $\omega \doteq\mathbb{E}[\delta] $. where $\delta_t$ is the TD error as follows:
\begin{equation}
\delta_t = r_{t+1}+\gamma
\theta_t^{\top}\phi_{t+1}-\theta_t^{\top}\phi_t.
\label{delta}
\end{equation}
Clearly, it is no longer to minimize Bellman errors.
First, the parameter $\omega$ is derived directly based on
stochastic gradient descent:
\begin{equation}
\omega_{t+1}\leftarrow \omega_{t}+\beta_t(\delta_t-\omega_t),
\label{omega}
\end{equation}
The gradient of the (3) with respect to $\theta$ is Then, based on stochastic semi-gradient descent, the update of
the parameter $\theta$ is as follows:
\begin{equation}
\theta_{t+1}\leftarrow
\theta_{t}+\alpha_t(\delta_t-\omega_t)\phi_t.
\label{theta}
\end{equation}
The semi-gradient of the (2) with respect to $\theta$ is
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
-\frac{1}{2}\nabla \text{VPBE}(\bm{\theta}) &=& -\mathbb{E}\Big[\Big( (\gamma \bm{\phi}' - \bm{\phi}) - \mathbb{E}[ (\gamma \bm{\phi}' - \bm{\phi})]\Big)\bm{\phi}^{\top} \Big] \\ &&-\frac{1}{2}\nabla \text{VBE}({\theta}) \\
& & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1} \mathbb{E}[( \delta -\mathbb{E}[ \delta])\bm{\phi}]\\ &=& \mathbb{E}[(\mathbb{E}[\delta_t|S_t]-\mathbb{E}[\mathbb{E}[\delta_t|S_t]])(\phi_t -\mathbb{E}[\phi_t])] \\
&=& \mathbb{E}\Big[\Big( (\bm{\phi} - \gamma \bm{\phi}')- \mathbb{E}[ (\bm{\phi} - \gamma \bm{\phi}')]\Big)\bm{\phi}^{\top} \Big] \\ &=& \mathbb{E}[\delta_t \phi_t] -\mathbb{E}[\delta_t] \mathbb{E}[\phi_t] ,
& & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\ % &=&-\mathbb{E}\Big[\Big( (\phi_t - \gamma\phi_t')- \mathbb{E}[ (\phi_t- \gamma {\phi_t}')]\Big)\phi_t^{\top} \Big]\theta + \mathbb{E}( r_{t+1}- \mathbb{E}[r_{t+1}])\bm{\phi_t}
& & \mathbb{E}\Big[\Big( r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta}\\ \end{array}
& & \hspace{2em} -\mathbb{E}[ r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta}]\Big)\bm{\phi} \Big]\\
&=& \textbf{A}^{\top} \textbf{C}^{-1}(-\textbf{A}\bm{\theta} + \textbf{b})
\end{array}
\end{equation*} \end{equation*}
where The key matrix $\textbf{A}_{\text{VMTD}}$ and $b_{\text{VMTD}}$ of on-policy VMTD is
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\textbf{A} &=& \mathbb{E}\Big[\Big( (\bm{\phi} - \gamma \bm{\phi}')- \mathbb{E}[ (\bm{\phi} - \gamma \bm{\phi}')]\Big)\bm{\phi}^{\top} \Big] \\ &&\textbf{A}_{\text{VMTD}} \\
&=& \mathbb{E}[(\bm{\phi} - \gamma \bm{\phi}')\bm{\phi}^{\top}] - \mathbb{E}[\bm{\phi} - \gamma \bm{\phi}']\mathbb{E}[\bm{\phi}^{\top}]\\ &=& \mathbb{E}[(\phi - \gamma \phi')\phi^{\top}] - \mathbb{E}[\phi - \gamma \phi']\mathbb{E}[\phi^{\top}]\\
&=& \mathrm{Cov}(\bm{\bm{\phi}},\bm{\bm{\phi}}-\gamma\bm{\bm{\phi}}'), &=&\sum_{s}d_{\pi}(s)\phi(s)\Big(\phi(s) -\gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\phi(s') \Big)^{\top} \\
&& -\sum_{s}d_{\pi}(s)\phi(s) \cdot \sum_{s}d_{\pi}(s)\Big(\phi(s) -\gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\phi(s') \Big)^{\top}\\
&=& \bm{\Phi}^{\top}\textbf{D}_{\mu}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi} -\bm{\Phi}^{\top}d_{\pi}d_{\pi}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}\\
&=& \bm{\Phi}^{\top}(\textbf{D}_{\pi}-d_{\pi}d_{\pi}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi},
\end{array} \end{array}
\end{equation*} \end{equation*}
% \begin{equation*}
% \begin{array}{ccl}
% \textbf{C} &=& \mathbb{E}[\bm{\bm{\phi}}\bm{\bm{\phi}}^{\top}],
% \end{array}
% \end{equation*}
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\textbf{C} &=& \mathbb{E}[\bm{\bm{\phi}}\bm{\bm{\phi}}^{\top}], &&b_{\text{VMTD}}\\
&=& \mathbb{E}( r- \mathbb{E}[r])\phi \\
&=& \mathbb{E}[r\phi] - \mathbb{E}[r]\mathbb{E}[\phi]\\
&=& \bm{\Phi}^{\top}(\textbf{D}_{\pi}-d_{\pi}d_{\pi}^{\top})r_\pi.
\end{array} \end{array}
\end{equation*} \end{equation*}
It can be easily obtained that The key matrix $\textbf{A}_{\text{VMTD}}$ and $b_{\text{VMTD}}$ of off-policy VMTD are, respectively,
\begin{equation*}
\textbf{A}_{\text{VMTD}} = \bm{\Phi}^{\top}(\textbf{D}_{\mu}-d_{\mu}d_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi},
\end{equation*}
\begin{equation*}
b_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}-d_{\mu}d_{\mu}^{\top})r_\pi,
\end{equation*}
In the on-policy 2-state environment, the minimum eigenvalue
of the key matrix for VMTD is greater than that of on-policy TDC and smaller than that of on-policy TD(0),
indicating that VMTD converges faster than TDC and slower than TD(0) in this
environment. In the off-policy 2-state environment, the
minimum eigenvalue of the key matrix for VMTD is greater than 0,
suggesting that VMTD can converge stably.
\subsection{Variance Minimization TDC Learning: VMTDC}
For off-policy learning, we propose a new objective function,
called Variance of Projected Bellman error (VPBE),
and the corresponding algorithm is called VMTDC.
\begin{align}
&\text{VPBE}(\bm{\theta}) \notag\\
&= \mathbb{E}[(\delta-\mathbb{E}[\delta]) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta ])\bm{\phi}] \\
% &= (\bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}))^{\top}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1} \notag \\
% & \bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
% &= (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}})^{\top}\textbf{D}^{\top}\bm{\Phi}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1} \notag \\
% & \bm{\Phi}^{\top}\textbf{D}(\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
% &= (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}})^{\top}\Pi^{\top}\textbf{D}\Pi \notag \\
% & (\textbf{W}_{\bm{\theta}} + \textbf{T}\textbf{V}_{\bm{\theta}} -\textbf{V}_{\bm{\theta}}) \notag \\
% &= (\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}-\textbf{W}_{\bm{\theta}}))^{\top}\textbf{D} \notag \\
% & (\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}-\textbf{W}_{\bm{\theta}})) \notag \\
% &= ||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}} - \textbf{W}_{\bm{\theta}})||^{2}_{\mu} \notag \\
% &= ||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}) - \Pi\textbf{W}_{\bm{\theta}}||^{2}_{\mu} \notag \\
&= \mathbb{E}[(\delta-\omega) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\omega)\bm{\phi}] ,
\end{align}
where
% $\textbf{W}_{\bm{\theta}}$ is viewed as vectors with every element being equal to
% $||\textbf{V}_{\bm{\theta}} - \textbf{T}\textbf{V}_{\bm{\theta}}||^{2}_{\mu}$ and
$\omega$ is used to approximate $\mathbb{E}[\delta]$, i.e., $\omega \doteq\mathbb{E}[\delta] $.
The gradient of the (6) with respect to $\theta$ is
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\textbf{b} &=& \mathbb{E}( r- \mathbb{E}[r])\bm{\phi} \\ -\frac{1}{2}\nabla \text{VPBE}({\theta}) &=& -\mathbb{E}\Big[\Big( (\gamma {\phi}' - {\phi}) - \mathbb{E}[ (\gamma {\phi}' - {\phi})]\Big){\phi}^{\top} \Big] \\
&=& \mathbb{E}[r\bm{\phi}] - \mathbb{E}[r]\mathbb{E}[\bm{\phi}]\\ & & \mathbb{E}[{\phi} {\phi}^{\top}]^{-1} \mathbb{E}[( \delta -\mathbb{E}[ \delta]){\phi}]\\
&=& \mathrm{Cov}(r,\bm{\bm{\phi}}), &=& \mathbb{E}\Big[\Big( ({\phi} - \gamma {\phi}')- \mathbb{E}[ ({\phi} - \gamma {\phi}')]\Big){\phi}^{\top} \Big] \\
& & \mathbb{E}[{\phi} {\phi}^{\top}]^{-1}\\
& & \mathbb{E}\Big[\Big( r + \gamma {{\phi}'}^{\top} {\theta} -{\phi}^{\top} {\theta}\\
& & \hspace{2em} -\mathbb{E}[ r + \gamma {{\phi}'}^{\top} {\theta} -{\phi}^{\top} {\theta}]\Big){\phi} \Big].
% &=& \textbf{A}^{\top} \textbf{C}^{-1}(-\textbf{A}\bm{\theta} + \textbf{b})
\end{array} \end{array}
\end{equation*} \end{equation*}
where $\mathrm{Cov}(\cdot,\cdot )$ is a covariance operator. It can be easily obtained that The key matrix $\textbf{A}_{\text{VMTDC}}$ and $b_{\text{VMTDC}}$ of VMTDC are, respectively,
\begin{equation*}
\textbf{A}_{\text{VMTDC}} = \textbf{A}_{\text{VMTD}}^{\top} \textbf{C}^{-1}\textbf{A}_{\text{VMTD}},
\end{equation*}
\begin{equation*}
b_{\text{VMTDC}}=\textbf{A}_{\text{VMTD}}^{\top} \textbf{C}^{-1}b_{\text{VMTD}},
\end{equation*}
where, for on-policy, $\textbf{A}_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\pi}-d_{\pi}d_{\pi}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$
and $b_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\pi}-d_{\pi}d_{\pi}^{\top})r_\pi$ and, for off-policy,
$\textbf{A}_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}-d_{\mu}d_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$
and $b_{\text{VMTD}}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}-d_{\mu}d_{\mu}^{\top})r_\pi$.
In the process of computing the gradient of the (4) with respect to $\theta$, In the process of computing the gradient of the (7) with respect to $\theta$,
$\omega$ is treated as a constant. $\omega$ is treated as a constant.
So, the derivation process of the VMTDC algorithm is the same So, the derivation process of the VMTDC algorithm is the same
as that of the TDC algorithm, the only difference is that the original $\delta$ is replaced by $\delta-\omega$. as that of the TDC algorithm, the only difference is that the original $\delta$ is replaced by $\delta-\omega$.
...@@ -133,60 +215,67 @@ Therefore, we can easily get the updated formula of VMTDC, as follows: ...@@ -133,60 +215,67 @@ Therefore, we can easily get the updated formula of VMTDC, as follows:
\end{equation} \end{equation}
and and
\begin{equation} \begin{equation}
\omega_{k+1}\leftarrow \omega_{k}+\beta_k (\delta_k- \omega_k), \omega_{k+1}\leftarrow \omega_{k}+\beta_k (\delta_k- \omega_k).
\label{omegavmtdc} \label{omegavmtdc}
\end{equation} \end{equation}
The VMTDC algorithm (\ref{thetavmtdc}) is derived to work The VMTDC algorithm (\ref{thetavmtdc}) is derived to work
with a given set of sub-samples—in the form of with a given set of sub-samples—in the form of
triples $(S_k, R_k, S'_k)$ that match transitions triples $(S_k, R_k, S'_k)$ that match transitions
from both the behavior and target policies. What if from both the behavior and target policies.
we wanted to use all the data? The data
is generated according to the behavior policy $\pi_b$,
while our objective is to learn about the target
policy $\pi$. We should use importance-sampling.
The VPBE with importance sampling is:
\begin{equation}
\label{rho_VPBE}
\begin{array}{ccl}
\text{VPBE}(\bm{\theta})&=&\mathbb{E}[(\rho\delta-\mathbb{E}[\rho\delta]) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\
& &\mathbb{E}[(\rho\delta -\mathbb{E}[\rho\delta ])\bm{\phi}],
\end{array}
\end{equation}
Following the linear VMTDC derivation, we get the following algorithm (linear VMTDC algorithm
based on importance weighting scenario):
\begin{equation}
\bm{\theta}_{k+1}\leftarrow\bm{\theta}_{k}+\alpha_{k}[(\rho_k\delta_{k}- \omega_k) \bm{\phi}_k\\
- \gamma\rho_k\bm{\phi}_{k+1}(\bm{\phi}^{\top}_k \bm{u}_{k})],
\end{equation}
\begin{equation}
\bm{u}_{k+1}\leftarrow \bm{u}_{k}+\zeta_{k}[(\rho_k\delta_{k}-\omega_k) - \bm{\phi}^{\top}_k \bm{u}_{k}]\bm{\phi}_k,
\end{equation}
and
\begin{equation}
\omega_{k+1}\leftarrow \omega_{k}+\beta_k (\rho_k\delta_k- \omega_k),
\end{equation}
The gradient of the (\ref{rho_VPBE}) with respect to $\theta$ is % What if
\begin{equation*} % we wanted to use all the data? The data
\begin{array}{ccl} % is generated according to the behavior policy $\pi_b$,
-\frac{1}{2}\nabla \text{VPBE}(\bm{\theta}) &=& \mathbb{E}\Big[\Big( \rho(\bm{\phi} - \gamma \bm{\phi}')- \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')]\Big)\bm{\phi}^{\top} \Big] \\ % while our objective is to learn about the target
& & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\ % policy $\pi$. We should use importance-sampling.
& & \mathbb{E}\Big[\Big( \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})\\ % The VPBE with importance sampling is:
& & \hspace{2em} -\mathbb{E}[ \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})]\Big)\bm{\phi} \Big]\\ % \begin{equation}
&=& \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')\bm{\phi}^{\top}]- \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')]\mathbb{E}[\bm{\phi}^{\top}] \\ % \label{rho_VPBE}
& & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\ % \begin{array}{ccl}
& & \mathbb{E}\Big[\Big( \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})\\ % \text{VPBE}(\bm{\theta})&=&\mathbb{E}[(\rho\delta-\mathbb{E}[\rho\delta]) \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\
& & \hspace{2em} -\mathbb{E}[ \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})]\Big)\bm{\phi} \Big]\\ % & &\mathbb{E}[(\rho\delta -\mathbb{E}[\rho\delta ])\bm{\phi}],
% &=&\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}\\ % \end{array}
&=& \textbf{A}^{\top} \textbf{C}^{-1}(-\textbf{A}\bm{\theta} + \textbf{b}), % \end{equation}
\end{array} % Following the linear VMTDC derivation, we get the following algorithm (linear VMTDC algorithm
\end{equation*} % based on importance weighting scenario):
where $\textbf{A}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$, % \begin{equation}
$\textbf{b}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})\textbf{r}_{\pi}$ and % \bm{\theta}_{k+1}\leftarrow\bm{\theta}_{k}+\alpha_{k}[(\rho_k\delta_{k}- \omega_k) \bm{\phi}_k\\
$\textbf{r}_{\pi}$ is viewed as vectors. % - \gamma\rho_k\bm{\phi}_{k+1}(\bm{\phi}^{\top}_k \bm{u}_{k})],
In the 2-state counterexample, % \end{equation}
$\textbf{A}_{\text{VMTDC}}=0.025$, meaning that VMTDC can stably converge and converges faster than TDC. % \begin{equation}
% \bm{u}_{k+1}\leftarrow \bm{u}_{k}+\zeta_{k}[(\rho_k\delta_{k}-\omega_k) - \bm{\phi}^{\top}_k \bm{u}_{k}]\bm{\phi}_k,
% \end{equation}
% and
% \begin{equation}
% \omega_{k+1}\leftarrow \omega_{k}+\beta_k (\rho_k\delta_k- \omega_k),
% \end{equation}
% The gradient of the (\ref{rho_VPBE}) with respect to $\theta$ is
% \begin{equation*}
% \begin{array}{ccl}
% -\frac{1}{2}\nabla \text{VPBE}(\bm{\theta}) &=& \mathbb{E}\Big[\Big( \rho(\bm{\phi} - \gamma \bm{\phi}')- \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')]\Big)\bm{\phi}^{\top} \Big] \\
% & & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\
% & & \mathbb{E}\Big[\Big( \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})\\
% & & \hspace{2em} -\mathbb{E}[ \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})]\Big)\bm{\phi} \Big]\\
% &=& \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')\bm{\phi}^{\top}]- \mathbb{E}[ \rho(\bm{\phi} - \gamma \bm{\phi}')]\mathbb{E}[\bm{\phi}^{\top}] \\
% & & \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\\
% & & \mathbb{E}\Big[\Big( \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})\\
% & & \hspace{2em} -\mathbb{E}[ \rho(r + \gamma {\bm{\phi}'}^{\top} \bm{\theta} -\bm{\phi}^{\top} \bm{\theta})]\Big)\bm{\phi} \Big]\\
% % &=&\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}\\
% &=& \textbf{A}^{\top} \textbf{C}^{-1}(-\textbf{A}\bm{\theta} + \textbf{b}),
% \end{array}
% \end{equation*}
% where $\textbf{A}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$,
% $\textbf{b}=\bm{\Phi}^{\top}(\textbf{D}_{\mu}- \textbf{d}_{\mu}\textbf{d}_{\mu}^{\top})\textbf{r}_{\pi}$ and
% $\textbf{r}_{\pi}$ is viewed as vectors.
In the on-policy 2-state environment, the minimum eigenvalue
of the key matrix for VMTDC is smaller than that of TD(0), TDC and VMTD
indicating that VMTDC converges slower than them in this
on-policy. In the off-policy 2-state environment, the
minimum eigenvalue of the key matrix for VMTD is greater than TDC,
suggesting that VMTDC converges faster than them in off-policy
environment.
...@@ -213,24 +302,24 @@ $\textbf{A}_{\text{VMTDC}}=0.025$, meaning that VMTDC can stably converge and co ...@@ -213,24 +302,24 @@ $\textbf{A}_{\text{VMTDC}}=0.025$, meaning that VMTDC can stably converge and co
% \label{deltaQ} % \label{deltaQ}
% \end{equation} % \end{equation}
% and $A^{*}_{k+1}={\arg \max}_{a}(\bm{\theta}_{k}^{\top}\bm{\phi}(s_{k+1},a))$. % and $A^{*}_{k+1}={\arg \max}_{a}(\bm{\theta}_{k}^{\top}\bm{\phi}(s_{k+1},a))$.
\begin{table*}[t] % \begin{table*}[t]
\caption{Minimum eigenvalues of various algorithms in the 2-state counterexample.} % \caption{Minimum eigenvalues of various algorithms in the 2-state counterexample.}
\vskip 0.15in % \vskip 0.15in
\begin{center} % \begin{center}
\begin{small} % \begin{small}
\begin{sc} % \begin{sc}
\begin{tabular}{lcccccr} % \begin{tabular}{lcccccr}
\toprule % \toprule
algorithm & off-policy TD & TDC & ETD & VMTDC & VMETD \\ % algorithm & off-policy TD & TDC & ETD & VMTDC & VMETD \\
\midrule % \midrule
Minimum eigenvalues&$-0.2$ & $0.016$ & $3.4$ & $0.025$ & $1.15$ \\ % Minimum eigenvalues&$-0.2$ & $0.016$ & $3.4$ & $0.025$ & $1.15$ \\
\bottomrule % \bottomrule
\end{tabular} % \end{tabular}
\end{sc} % \end{sc}
\end{small} % \end{small}
\end{center} % \end{center}
\vskip -0.1in % \vskip -0.1in
\end{table*} % \end{table*}
\subsection{Variance Minimization ETD Learning: VMETD} \subsection{Variance Minimization ETD Learning: VMETD}
Based on the off-policy TD algorithm, a scalar, $F$, Based on the off-policy TD algorithm, a scalar, $F$,
is introduced to obtain the ETD algorithm, is introduced to obtain the ETD algorithm,
...@@ -242,65 +331,73 @@ VMETD by the following update: ...@@ -242,65 +331,73 @@ VMETD by the following update:
% \delta_{t}= R_{t+1}+\gamma \theta_t^{\top}\phi_{t+1}-\theta_t^{\top}\phi_t. % \delta_{t}= R_{t+1}+\gamma \theta_t^{\top}\phi_{t+1}-\theta_t^{\top}\phi_t.
% \end{equation} % \end{equation}
\begin{equation} \begin{equation}
\label{fvmetd}
F_t \leftarrow \gamma \rho_{t-1}F_{t-1}+1,
\end{equation}
\begin{equation}
\label{thetavmetd} \label{thetavmetd}
\bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\alpha_k (F_k \rho_k\delta_k - \omega_{k})\bm{\phi}_k, {\theta}_{t+1}\leftarrow {\theta}_t+\alpha_t (F_t \rho_t\delta_t - \omega_{t}){\phi}_t,
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\label{omegavmetd} \label{omegavmetd}
\omega_{k+1} \leftarrow \omega_k+\beta_k(F_k \rho_k \delta_k - \omega_k), \omega_{t+1} \leftarrow \omega_t+\beta_t(F_t \rho_t \delta_t - \omega_t),
\end{equation} \end{equation}
where $\omega$ is used to estimate $\mathbb{E}[F \rho\delta]$, i.e., $\omega \doteq \mathbb{E}[F \rho\delta]$. where $\omega$ is used to estimate $\mathbb{E}[F \rho\delta]$, i.e., $\omega \doteq \mathbb{E}[F \rho\delta]$.
(\ref{thetavmetd}) can be rewritten as (\ref{thetavmetd}) can be rewritten as
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\bm{\theta}_{k+1}&\leftarrow& \bm{\theta}_k+\alpha_k (F_k \rho_k\delta_k - \omega_k)\bm{\phi}_k -\alpha_k \omega_{k+1}\bm{\phi}_k\\ {\theta}_{t+1}&\leftarrow& {\theta}_t+\alpha_t (F_t \rho_t\delta_t - \omega_t){\phi}_t -\alpha_t \omega_{t+1}{\phi}_t\\
&=&\bm{\theta}_{k}+\alpha_k(F_k\rho_k\delta_k-\mathbb{E}_{\mu}[F_k\rho_k\delta_k|\bm{\theta}_k])\bm{\phi}_k\\ &=&{\theta}_{t}+\alpha_t(F_t\rho_t\delta_t-\mathbb{E}_{\mu}[F_t\rho_t\delta_t|{\theta}_t]){\phi}_t\\
&=&\bm{\theta}_k+\alpha_k F_k \rho_k (R_{k+1}+\gamma \bm{\theta}_k^{\top}\bm{\phi}_{k+1}-\bm{\theta}_k^{\top}\bm{\phi}_k)\bm{\phi}_k\\ &=&{\theta}_t+\alpha_t F_t \rho_t (r_{t+1}+\gamma {\theta}_t^{\top}{\phi}_{t+1}-{\theta}_t^{\top}{\phi}_t){\phi}_t\\
& & \hspace{2em} -\alpha_k \mathbb{E}_{\mu}[F_k \rho_k \delta_k]\bm{\phi}_k\\ & & \hspace{2em} -\alpha_t \mathbb{E}_{\mu}[F_t \rho_t \delta_t]{\phi}_t\\
&=& \bm{\theta}_k+\alpha_k \{\underbrace{(F_k\rho_kR_{k+1}-\mathbb{E}_{\mu}[F_k\rho_k R_{k+1}])\bm{\phi}_k}_{\textbf{b}_{\text{VMETD},k}}\\ &=& {\theta}_t+\alpha_t \{\underbrace{(F_t\rho_tr_{t+1}-\mathbb{E}_{\mu}[F_t\rho_t r_{t+1}]){\phi}_t}_{{b}_{\text{VMETD},t}}\\
&&\hspace{-5em}- \underbrace{(F_k\rho_k\bm{\phi}_k(\bm{\phi}_k-\gamma\bm{\phi}_{k+1})^{\top}-\bm{\phi}_k\mathbb{E}_{\mu}[F_k\rho_k (\bm{\phi}_k-\gamma\bm{\phi}_{k+1})]^{\top})}_{\textbf{A}_{\text{VMETD},k}}\bm{\theta}_k\}. &&\hspace{-7em}- \underbrace{(F_t\rho_t{\phi}_t({\phi}_t-\gamma{\phi}_{t+1})^{\top}-{\phi}_t\mathbb{E}_{\mu}[F_t\rho_t ({\phi}_t-\gamma{\phi}_{t+1})]^{\top})}_{\textbf{A}_{\text{VMETD},t}}{\theta}_t\}.
\end{array} \end{array}
\end{equation*} \end{equation*}
Therefore, Therefore,
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\textbf{A}_{\text{VMETD}}&=&\lim_{k \rightarrow \infty} \mathbb{E}[\textbf{A}_{\text{VMETD},k}]\\ &&\textbf{A}_{\text{VMETD}}\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k \rho_k \bm{\phi}_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]\\ &=&\lim_{t \rightarrow \infty} \mathbb{E}[\textbf{A}_{\text{VMETD},t}]\\
&&\hspace{-1em}- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[F_t \rho_t {\phi}_t ({\phi}_t - \gamma {\phi}_{t+1})^{\top}]\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]\\ &&\hspace{1em}- \lim_{t\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_t]\mathbb{E}_{\mu}[F_t \rho_t ({\phi}_t - \gamma {\phi}_{t+1})]^{\top}\\
&&\hspace{-1em}- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ % &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_tF_t \rho_t ({\phi}_t - \gamma {\phi}_{t+1})^{\top}]\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})^{\top}]\\ % &&\hspace{1em}- \lim_{t\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_t]\mathbb{E}_{\mu}[F_t \rho_t ({\phi}_t - \gamma {\phi}_{t+1})]^{\top}\\
&&\hspace{-2em}-\lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k \rho_k (\bm{\phi}_k - \gamma \bm{\phi}_{k+1})]^{\top}\\ &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_tF_t \rho_t ({\phi}_t - \gamma {\phi}_{t+1})^{\top}]\\
&& \hspace{-9em}=\sum_{s} d_{\mu}(s)\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k|S_k = s]\mathbb{E}_{\mu}[\rho_k\phi_k(\phi_k - \gamma \phi_{k+1})^{\top}|S_k = s]\\ &&\hspace{1em}-\lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_t]\lim_{t \rightarrow \infty}\mathbb{E}_{\mu}[F_t \rho_t ({\phi}_t - \gamma {\phi}_{t+1})]^{\top}\\
&&\hspace{-2em}-\sum_{s} d_{\mu}(s)\phi(s)\sum_{s} d_{\mu}(s)\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k|S_k = s]\\ && \hspace{-2em}=\sum_{s} d_{\mu}(s)\lim_{t \rightarrow \infty}\mathbb{E}_{\mu}[F_t|S_t = s]\mathbb{E}_{\mu}[\rho_t\phi_t(\phi_t - \gamma \phi_{t+1})^{\top}|S_t= s]\\
&&\hspace{2em}\mathbb{E}_{\mu}[\rho_k(\phi_k - \gamma \phi_{k+1})^{\top}|S_k = s]\\ &&\hspace{1em}-\sum_{s} d_{\mu}(s)\phi(s)\sum_{s} d_{\mu}(s)\lim_{t \rightarrow \infty}\mathbb{E}_{\mu}[F_t|S_t = s]\\
&=& \sum_{s} f(s)\mathbb{E}_{\pi}[\phi_k(\phi_k - \gamma \phi_{k+1})^{\top}|S_k = s]\\ &&\hspace{7em}\mathbb{E}_{\mu}[\rho_t(\phi_t - \gamma \phi_{t+1})^{\top}|S_t = s]\\
&&\hspace{-3em}-\sum_{s} d_{\mu}(s)\phi(s)\sum_{s} f(s)\mathbb{E}_{\pi}[(\phi_k - \gamma \phi_{k+1})^{\top}|S_k = s]\\ &=& \sum_{s} f(s)\mathbb{E}_{\pi}[\phi_t(\phi_t- \gamma \phi_{t+1})^{\top}|S_t = s]\\
&&\hspace{1em}-\sum_{s} d_{\mu}(s)\phi(s)\sum_{s} f(s)\mathbb{E}_{\pi}[(\phi_t- \gamma \phi_{t+1})^{\top}|S_t = s]\\
&=&\sum_{s} f(s) \bm{\phi}(s)(\bm{\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\bm{\phi}(s'))^{\top} \\ &=&\sum_{s} f(s) \bm{\phi}(s)(\bm{\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\bm{\phi}(s'))^{\top} \\
&&\hspace{-4em} -\sum_{s} d_{\mu}(s) \bm{\phi}(s) * \sum_{s} f(s)(\bm{\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}\bm{\phi}(s'))^{\top}\\ &&\hspace{1em} -\sum_{s} d_{\mu}(s) {\phi}(s) * \sum_{s} f(s)({\phi}(s) - \gamma \sum_{s'}[\textbf{P}_{\pi}]_{ss'}{\phi}(s'))^{\top}\\
&=&{\bm{\Phi}}^{\top} \textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}) \bm{\Phi} - {\bm{\Phi}}^{\top} \textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\mu}) \bm{\Phi} \\ &=&{\bm{\Phi}}^{\top} \textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi}) \bm{\Phi} - {\bm{\Phi}}^{\top} {d}_{\mu} {f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\mu}) \bm{\Phi} \\
&=&{\bm{\Phi}}^{\top} (\textbf{F} - \textbf{d}_{\mu} \textbf{f}^{\top}) (\textbf{I} - \gamma \textbf{P}_{\pi}){\bm{\Phi}} \\ &=&{\bm{\Phi}}^{\top} (\textbf{F} - {d}_{\mu} {f}^{\top}) (\textbf{I} - \gamma \textbf{P}_{\pi}){\bm{\Phi}} \\
&=&{\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\pi})){\bm{\Phi}} \\ &=&{\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-{d}_{\mu} {f}^{\top} (\textbf{I} - \gamma \textbf{P}_{\pi})){\bm{\Phi}} \\
&=&{\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){\bm{\Phi}}, &=&{\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-{d}_{\mu} {d}_{\mu}^{\top} ){\bm{\Phi}},
\end{array} \end{array}
\end{equation*} \end{equation*}
\begin{equation*} \begin{equation*}
\begin{array}{ccl} \begin{array}{ccl}
\textbf{b}_{\text{VMETD}}&=&\lim_{k \rightarrow \infty} \mathbb{E}[\textbf{b}_{\text{VMETD},k}]\\ &&{b}_{\text{VMETD}}\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[F_k\rho_kR_{k+1}\bm{\phi}_k]\\ &=&\lim_{t \rightarrow \infty} \mathbb{E}[{b}_{\text{VMETD},t}]\\
&&- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[F_t\rho_tR_{t+1}{\phi}_t]\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k\rho_kR_{k+1}]\\ &&\hspace{2em} - \lim_{t\rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_t]\mathbb{E}_{\mu}[F_t\rho_kR_{k+1}]\\
&&- \lim_{k\rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\mathbb{E}_{\mu}[\bm{\phi}_k]\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_tF_t\rho_tr_{t+1}]\\
&=& \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[\bm{\phi}_kF_k\rho_kR_{k+1}]\\ &&\hspace{2em} - \lim_{t\rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_t]\mathbb{E}_{\mu}[{\phi}_t]\mathbb{E}_{\mu}[F_t\rho_tr_{t+1}]\\
&&- \lim_{k \rightarrow \infty} \mathbb{E}_{\mu}[ \bm{\phi}_k]\lim_{k \rightarrow \infty}\mathbb{E}_{\mu}[F_k\rho_kR_{k+1}]\\ &=& \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[{\phi}_tF_t\rho_tr_{t+1}]\\
&=&\sum_{s} f(s) \bm{\phi}(s)r_{\pi} - \sum_{s} d_{\mu}(s) \bm{\phi}(s) * \sum_{s} f(s)r_{\pi} \\ &&\hspace{2em} - \lim_{t \rightarrow \infty} \mathbb{E}_{\mu}[ {\phi}_t]\lim_{t \rightarrow \infty}\mathbb{E}_{\mu}[F_t\rho_tr_{t+1}]\\
&=&\bm{\bm{\Phi}}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi}. &=&\sum_{s} f(s) {\phi}(s)r_{\pi} - \sum_{s} d_{\mu}(s) {\phi}(s) * \sum_{s} f(s)r_{\pi} \\
&=&\bm{\bm{\Phi}}^{\top}(\textbf{F}-{d}_{\mu} {f}^{\top}){r}_{\pi}.
\end{array} \end{array}
\end{equation*} \end{equation*}
Therefore, in the 2-state counterexample, In the off-policy 2-state environment, the minimum eigenvalue
$\textbf{A}_{\text{VMETD}}=1.15$, meaning that VMETD can stably converge and converges slower than ETD. of the key matrix for VMETD is greater than that of TD(0), TDC and VMTD and smaller than that of ETD,
However, subsequent experiments showed that the VMETD algorithm converges more smoothly and performs better in controlled experiments. indicating that VMTDC converges faster than TD(0), TDC and VMTD and slower than ETD in this
off-policy.
However, subsequent experiments showed that the VMETD algorithm converges more smoothly and performs best in controlled experiments.
% In this paper, we refer to the control algorithm for ETD as EQ. % In this paper, we refer to the control algorithm for ETD as EQ.
% Now, we will introduce the improved version of the EQ algorithm, named VMEQ: % Now, we will introduce the improved version of the EQ algorithm, named VMEQ:
% \begin{equation} % \begin{equation}
......
NEW_aaai/main/pic/BairdExample copy 2.tex 0 → 100644
\resizebox{7cm}{4.4cm}{
\begin{tikzpicture}[smooth]
\node[coordinate] (origin) at (0.3,0) {};
\node[coordinate] (num7) at (3,0) {};
\node[coordinate] (num1) at (1,2.5) {};
\path (num7) ++ (-10:0.5cm) node (num7_bright1) [coordinate] {};
\path (num7) ++ (-30:0.7cm) node (num7_bright2) [coordinate] {};
\path (num7) ++ (-60:0.35cm) node (num7_bright3) [coordinate] {};
\path (num7) ++ (-60:0.6cm) node (num7_bright4) [coordinate] {};
\path (origin) ++ (90:3cm) node (origin_above) [coordinate] {};
\path (origin_above) ++ (0:5.7cm) node (origin_aright) [coordinate] {};
\path (num1) ++ (90:0.5cm) node (num1_a) [coordinate] {};
\path (num1) ++ (-90:0.3cm) node (num1_b) [coordinate] {};
\path (num1) ++ (0:1cm) node (num2) [coordinate] {};
\path (num1_a) ++ (0:1cm) node (num2_a) [coordinate] {};
\path (num1_b) ++ (0:1cm) node (num2_b) [coordinate] {};
\path (num2) ++ (0:1cm) node (num3) [coordinate] {};
\path (num2_a) ++ (0:1cm) node (num3_a) [coordinate] {};
\path (num2_b) ++ (0:1cm) node (num3_b) [coordinate] {};
\path (num3) ++ (0:1cm) node (num4) [coordinate] {};
\path (num3_a) ++ (0:1cm) node (num4_a) [coordinate] {};
\path (num3_b) ++ (0:1cm) node (num4_b) [coordinate] {};
\path (num4) ++ (0:1cm) node (num5) [coordinate] {};
\path (num4_a) ++ (0:1cm) node (num5_a) [coordinate] {};
\path (num4_b) ++ (0:1cm) node (num5_b) [coordinate] {};
\path (num5) ++ (0:1cm) node (num6) [coordinate] {};
\path (num5_a) ++ (0:1cm) node (num6_a) [coordinate] {};
\path (num5_b) ++ (0:1cm) node (num6_b) [coordinate] {};
%\draw[->](0,0) -- (1,1);
%\draw[dashed,line width = 0.03cm] (0,0) -- (1,1);
%\fill (0.5,0.5) circle (0.5);
%\draw[shape=circle,fill=white,draw=black] (a) at (num7) {7};
\draw[dashed,line width = 0.03cm,xshift=3cm] plot[tension=0.06]
coordinates{(num7) (origin) (origin_above) (origin_aright)};
\draw[->,>=stealth,line width = 0.02cm,xshift=3cm] plot[tension=0.5]
coordinates{(num7) (num7_bright1) (num7_bright2)(num7_bright4) (num7_bright3)};
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (g) at (num7) {7};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num1) -- (num1_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (a) at (num1_b) {1};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num2) -- (num2_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (b) at (num2_b) {2};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num3) -- (num3_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (c) at (num3_b) {3};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num4) -- (num4_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (d) at (num4_b) {4};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num5) -- (num5_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (e) at (num5_b) {5};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num6) -- (num6_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (f) at (num6_b) {6};
\draw[->,>=stealth,line width = 0.02cm] (a)--(g);
\draw[->,>=stealth,line width = 0.02cm] (b)--(g);
\draw[->,>=stealth,line width = 0.02cm] (c)--(g);
\draw[->,>=stealth,line width = 0.02cm] (d)--(g);
\draw[->,>=stealth,line width = 0.02cm] (e)--(g);
\draw[->,>=stealth,line width = 0.02cm] (f)--(g);
\end{tikzpicture}
}
NEW_aaai/main/pic/BairdExample copy.tex 0 → 100644
\resizebox{7cm}{4.4cm}{
\begin{tikzpicture}[smooth]
\node[coordinate] (origin) at (0.3,0) {};
\node[coordinate] (num7) at (3,0) {};
\node[coordinate] (num1) at (1,2.5) {};
\path (num7) ++ (-10:0.5cm) node (num7_bright1) [coordinate] {};
\path (num7) ++ (-30:0.7cm) node (num7_bright2) [coordinate] {};
\path (num7) ++ (-60:0.35cm) node (num7_bright3) [coordinate] {};
\path (num7) ++ (-60:0.6cm) node (num7_bright4) [coordinate] {};
\path (origin) ++ (90:3cm) node (origin_above) [coordinate] {};
\path (origin_above) ++ (0:5.7cm) node (origin_aright) [coordinate] {};
\path (num1) ++ (90:0.5cm) node (num1_a) [coordinate] {};
\path (num1) ++ (-90:0.3cm) node (num1_b) [coordinate] {};
\path (num1) ++ (0:1cm) node (num2) [coordinate] {};
\path (num1_a) ++ (0:1cm) node (num2_a) [coordinate] {};
\path (num1_b) ++ (0:1cm) node (num2_b) [coordinate] {};
\path (num2) ++ (0:1cm) node (num3) [coordinate] {};
\path (num2_a) ++ (0:1cm) node (num3_a) [coordinate] {};
\path (num2_b) ++ (0:1cm) node (num3_b) [coordinate] {};
\path (num3) ++ (0:1cm) node (num4) [coordinate] {};
\path (num3_a) ++ (0:1cm) node (num4_a) [coordinate] {};
\path (num3_b) ++ (0:1cm) node (num4_b) [coordinate] {};
\path (num4) ++ (0:1cm) node (num5) [coordinate] {};
\path (num4_a) ++ (0:1cm) node (num5_a) [coordinate] {};
\path (num4_b) ++ (0:1cm) node (num5_b) [coordinate] {};
\path (num5) ++ (0:1cm) node (num6) [coordinate] {};
\path (num5_a) ++ (0:1cm) node (num6_a) [coordinate] {};
\path (num5_b) ++ (0:1cm) node (num6_b) [coordinate] {};
%\draw[->](0,0) -- (1,1);
%\draw[dashed,line width = 0.03cm] (0,0) -- (1,1);
%\fill (0.5,0.5) circle (0.5);
%\draw[shape=circle,fill=white,draw=black] (a) at (num7) {7};
\draw[dashed,line width = 0.03cm,xshift=3cm] plot[tension=0.06]
coordinates{(num7) (origin) (origin_above) (origin_aright)};
\draw[->,>=stealth,line width = 0.02cm,xshift=3cm] plot[tension=0.5]
coordinates{(num7) (num7_bright1) (num7_bright2)(num7_bright4) (num7_bright3)};
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (g) at (num7) {7};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num1) -- (num1_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (a) at (num1_b) {1};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num2) -- (num2_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (b) at (num2_b) {2};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num3) -- (num3_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (c) at (num3_b) {3};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num4) -- (num4_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (d) at (num4_b) {4};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num5) -- (num5_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (e) at (num5_b) {5};
\draw[<->,>=stealth,dashed,line width = 0.03cm,] (num6) -- (num6_a) ;
\node[line width = 0.02cm,shape=circle,fill=white,draw=black] (f) at (num6_b) {6};
\draw[->,>=stealth,line width = 0.02cm] (a)--(g);
\draw[->,>=stealth,line width = 0.02cm] (b)--(g);
\draw[->,>=stealth,line width = 0.02cm] (c)--(g);
\draw[->,>=stealth,line width = 0.02cm] (d)--(g);
\draw[->,>=stealth,line width = 0.02cm] (e)--(g);
\draw[->,>=stealth,line width = 0.02cm] (f)--(g);
\end{tikzpicture}
}
NEW_aaai/main/pic/cl.pdf
No preview for this file type
NEW_aaai/main/pic/mt.pdf
No preview for this file type
NEW_aaai/main/preliminaries.tex
...@@ -19,172 +19,243 @@ Reinforcement learning agent interacts with environment, observes state, ...@@ -19,172 +19,243 @@ Reinforcement learning agent interacts with environment, observes state,
A policy is a mapping $\pi:S\times A \rightarrow [0,1]$. The goal of the A policy is a mapping $\pi:S\times A \rightarrow [0,1]$. The goal of the
agent is to find an optimal policy $\pi^*$ to maximize the expectation of a agent is to find an optimal policy $\pi^*$ to maximize the expectation of a
discounted cumulative rewards in a long period. discounted cumulative rewards in a long period. For each discrete time step
$t=0,1,2,3,…$,
State value function $V^{\pi}(s)$ for a stationary policy $\pi$ is State value function $V^{\pi}(s)$ for a stationary policy $\pi$ is
defined as: defined as:
\begin{equation*} \begin{equation*}
V^{\pi}(s)=\mathbb{E}_{\pi}[\sum_{k=0}^{\infty} \gamma^k R_{k}|s_0=s]. V^{\pi}(s)=\mathbb{E}_{\pi}[\sum_{k=0}^{\infty} \gamma^k R_{t+k+1}|S_t=s].
\label{valuefunction} \label{valuefunction}
\end{equation*} \end{equation*}
Linear value function for state $s\in S$ is defined as: Linear value function for state $s\in S$ is defined as:
\begin{equation} \begin{equation}
V_{{\theta}}(s):= {\bm{\theta}}^{\top}{\bm{\phi}}(s) = \sum_{i=1}^{m} V_{{\theta}}(s):= {{\theta}}^{\top}{{\phi}}(s) = \sum_{i=1}^{m}
\theta_i \phi_i(s), \theta_i \phi_i(s),
\label{linearvaluefunction} \label{linearvaluefunction}
\end{equation} \end{equation}
where ${\bm{\theta}}:=(\theta_1,\theta_2,\ldots,\theta_m)^{\top}\in where ${{\theta}}:=(\theta_1,\theta_2,\ldots,\theta_m)^{\top}\in
\mathbb{R}^m$ is a parameter vector, \mathbb{R}^m$ is a parameter vector,
${\bm{\phi}}:=(\phi_1,\phi_2,\ldots,\phi_m)^{\top}\in \mathbb{R}^m$ is a feature ${{\phi}}:=(\phi_1,\phi_2,\ldots,\phi_m)^{\top}\in \mathbb{R}^m$ is a feature
function defined on state space $S$, and $m$ is the feature size. function defined on state space $S$, and $m$ is the feature size.
Tabular temporal difference (TD) learning \cite{Sutton2018book} has been successfully applied to small-scale problems. Tabular temporal difference (TD) learning \cite{Sutton2018book} has been successfully applied to small-scale problems.
To deal with the well-known curse of dimensionality of large scale MDPs, value To deal with the well-known curse of dimensionality of large scale MDPs, value
function is usually approximated by a linear model (the focus of this paper), kernel methods, decision function is usually approximated by a linear model (the focus of this paper), kernel methods, decision
trees, or neural networks, etc. trees, or neural networks, etc.
% This paper focuses on the linear model.
% TD learning can also be used to find optimal strategies. The problem of finding an optimal policy is
% often called the control problem. Two popular TD methods are Sarsa and Q-leaning. The former is an on-policy
% TD control, while the latter is an off-policy control.
% It is well known that TDC algorithm \cite{sutton2009fast} guarantees
% convergence under off-policy conditions while the off-policy TD algorithm may diverge. The
% objective function of TDC is MSPBE.
% TDC is essentially an adjustment or correction of the TD update so that it
% follows the gradient of the MSPBE objective function. In the context of the TDC algorithm, the control algorithm
% is known as Greedy-GQ($\lambda$) \cite{sutton2009fast}. When $\lambda$ is set to 0, it is denoted
% as GQ(0).
\subsection{On-policy and Off-policy} \subsection{On-policy and Off-policy}
\begin{table*}[t]
\caption{Minimum eigenvalues of various algorithms in the 2-state counterexample.}
\label{tab:min_eigenvalues} % 添加标签
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lccccccr}
\toprule
algorithm &TD & TDC & ETD & VMTD & VMTDC & VMETD \\
\midrule
on-policy 2-state&$0.475$ & $0.09025$ & \text{\textbackslash}& $0.25$ & $0.025$ & \text{\textbackslash} \\
off-policy 2-state&$-0.2$ & $0.016$ & $3.4$ & $0.25$ & $0.025$ & $1.15$\\
\bottomrule
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table*}
On-policy and off-policy algorithms are currently hot topics in research. On-policy and off-policy algorithms are currently hot topics in research.
Off-policy algorithms, in particular, present greater challenges due to the
difficulty in ensuring their convergence, making them more complex to study.
The main difference between the two lies in the fact that in on-policy algorithms, The main difference between the two lies in the fact that in on-policy algorithms,
the behavior policy $\mu$ and the target policy $\pi$ are the same during the learning process. the behavior policy $\mu$ and the target policy $\pi$ are the same during the learning process.
The algorithm directly generates data from the current policy and optimizes it.
In off-policy algorithms, however, the behavior policy and the target policy are different. In off-policy algorithms, however, the behavior policy and the target policy are different.
The algorithm uses data generated from the behavior policy to optimize the The algorithm uses data generated from the behavior policy to optimize the
target policy, which leads to higher sample efficiency and complex stability issues. target policy, which leads to higher sample efficiency and complex stability issues.
Taking the TD(0) algorithm as an example can help understand the different % In the on-policy TD(0) algorithm, since the behavior policy and the target policy
performances of on-policy and off-policy: % are consistent, the convergence of TD(0) is more assured. In each time step $t$ of the
% update, the algorithm is based on the actual behavior of the current policy,
% which gradually leads the value function estimate to converge to the true
% value of the target policy.
In the on-policy TD(0) algorithm, the behavior policy and the target policy From the theory of stochastic methods, the
are the same. The algorithm uses the data generated by the current policy to convergence point of linear TD algorithms, is a parameter vector, say $\bm{\theta}$, that satisfies
update its value estimates. Since the behavior policy and the target policy
are consistent, the convergence of TD(0) is more assured. In each step of the
update, the algorithm is based on the actual behavior of the current policy,
which gradually leads the value function estimate to converge to the true
value of the target policy.
The on-policy TD(0) update formula is
\begin{equation*} \begin{equation*}
\label{thetatd_onpolicy}
\begin{array}{ccl} \begin{array}{ccl}
\bm{\theta}_{k+1}&\leftarrow&\bm{\theta}_k+\alpha_k \delta_k\bm{\phi}_k, b - \textbf{A}{\theta}&=&0,
\end{array} \end{array}
\end{equation*} \end{equation*}
where $\delta_k = r_{k+1}+\gamma \bm{\theta}_k^{\top}\bm{\phi}_{k+1}-\bm{\theta}_k^{\top}\bm{\phi}_k$ and where $\textbf{A}\in \mathbb{R}^{|S| \times m}$ and $b\in \mathbb{R}^{m}$.
the key matrix $\textbf{A}_{\text{on}}$ of on-policy TD(0) is If the matrix
$\textbf{A}$ is positive definite, then the algorithm converges.
The convergence rate of the algorithm is related to the matrix
$\textbf{A}$. The larger the minimum eigenvalue of
$\textbf{A}$, the faster the convergence rate.
Next, we will compute the minimum eigenvalue of
$\textbf{A}$ for TD(0), TDC, and ETD in both on-policy and off-policy settings in a 2-state environment.
First, we will introduce the environment setup for the 2-state case in both on-policy and off-policy settings.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.3\columnwidth, height=0.15\columnwidth]{main/pic/2StateExample.pdf}
\caption{2-state}
\end{center}
\end{figure}
The "1"$\rightarrow$"2" problem has only two states. From each
state, there are two actions, left and right, which take
the agent to the left or right state. All rewards are zero.
The feature $\bm{\Phi}=(1,2)^{\top}$
are assigned to the left and
the right state. The first policy takes the equal
probability to left or right
in both states, i.e.,
$
\textbf{P}_{1}=
\begin{bmatrix}
0.5 & 0.5 \\
0.5 & 0.5
\end{bmatrix}
$.
The second policy only selects action right in both states, i.e.,
$
\textbf{P}_{2}=
\begin{bmatrix}
0 & 1 \\
0 & 1
\end{bmatrix}
$.
The state distribution of
the first policy is $d_1 =(0.5,0.5)^{\top}$.
The state distribution of
the second policy is $d_1 =(0,1)^{\top}$.
The discount factor is $\gamma=0.9$.
In the on-policy setting, the behavior policy
and the target policy are the same, so
let $\textbf{P}_{\mu}=\textbf{P}_{\pi}=\textbf{P}_{1}$.
In the off-policy setting,
let $\textbf{P}_{\mu}=\textbf{P}_{1}$ and $\textbf{P}_{\pi}=\textbf{P}_{2}$.
% The on-policy TD(0) update formula is
% \begin{equation*}
% \label{thetatd_onpolicy}
% \begin{array}{ccl}
% \bm{\theta}_{t+1}&\leftarrow&\bm{\theta}_t+\alpha_t \delta_t\bm{\phi}_t,
% \end{array}
% \end{equation*}
% where $\delta_t = r_{t+1}+\gamma \bm{\theta}_t^{\top}\bm{\phi}_{t+1}-\bm{\theta}_t^{\top}\bm{\phi}_t$ is one-step TD error and
The key matrix $\textbf{A}_{\text{on}}$ of on-policy TD(0) is
\begin{equation*} \begin{equation*}
\textbf{A}_{\text{on}} = \bm{\Phi}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}, \textbf{A}_{\text{on}} = \bm{\Phi}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi},
\end{equation*} \end{equation*}
where $\bm{\Phi}$ is the $N \times n$ matrix with the $\phi(s)$ as its rows, and $\textbf{D}_{\pi}$ is the $N \times N$ diagonal where $\bm{\Phi}$ is the $|S| \times m$ matrix with the $\phi(s)$ as its rows, and $\textbf{D}_{\pi}$ is the $|S| \times |S|$ diagonal
matrix with $\textbf{d}_{\pi}$ on its diagonal. $\textbf{d}_{\pi}$ is a vector, each component representing the steady-state matrix with $d_{\pi}$ on its diagonal. $d_{\pi}$ is a vector, each component representing the steady-state
distribution under $\pi$. $\textbf{P}_{\pi}$ denote the $N \times N$ matrix of transition probabilities under $\pi$. And $\textbf{P}_{\pi}^{\top}\textbf{d}_{\pi}=\textbf{d}_{\pi}$. distribution under policy $\pi$. $\textbf{P}_{\pi}$ denote the $|S| \times |S|$ matrix of transition probabilities under $\pi$. And $\textbf{P}_{\pi}^{\top}d_{\pi}=d_{\pi}$.
An $\bm{\Phi}^{\top}\bm{\text{X}}\bm{\Phi}$ matrix of this
form will be positive definite whenever the matrix $\bm{\text{X}}$ is positive definite.
Any matrix $\bm{\text{X}}$ is positive definite if and only if
the symmetric matrix $\bm{\text{S}}=\bm{\text{X}}+\bm{\text{X}}^{\top}$ is positive definite.
Any symmetric real matrix $\bm{\text{S}}$ is positive definite if the absolute values of
its diagonal entries are greater than the sum of the absolute values of the corresponding
off-diagonal entries\cite{sutton2016emphatic}.
All components of the matrix $\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are positive.
The row sums of $\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are positive. And The row sums of
$\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are
\begin{equation*}
\begin{array}{ccl}
\textbf{1}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})&=&\textbf{d}_{\pi}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
&=& \textbf{d}_{\pi}^{\top} - \gamma \textbf{d}_{\pi}^{\top}\textbf{P}_{\pi}\\
&=& \textbf{d}_{\pi}^{\top} - \gamma \textbf{d}_{\pi}^{\top}\\
&=& (1-\gamma)\textbf{d}_{\pi}^{\top},
\end{array}
\end{equation*}
all components of which are positive. Thus, the key matrix and its $\textbf{A}_{\text{on}}$ matrix are positive
definite, and on-policy TD(0) is stable
The off-policy TD(0) update formula is % An $\bm{\Phi}^{\top}\bm{\text{X}}\bm{\Phi}$ matrix of this
\begin{equation*} % form will be positive definite whenever the matrix $\bm{\text{X}}$ is positive definite.
\label{thetatd_offpolicy} % Any matrix $\bm{\text{X}}$ is positive definite if and only if
\begin{array}{ccl} % the symmetric matrix $\bm{\text{S}}=\bm{\text{X}}+\bm{\text{X}}^{\top}$ is positive definite.
\bm{\theta}_{k+1}&\leftarrow&\bm{\theta}_k+\alpha_k \rho_k \delta_k\bm{\phi}_k, % Any symmetric real matrix $\bm{\text{S}}$ is positive definite if the absolute values of
\end{array} % its diagonal entries are greater than the sum of the absolute values of the corresponding
\end{equation*} % off-diagonal entries\cite{sutton2016emphatic}.
where $\rho_k =\frac{\pi(A_k | S_k)}{\mu(A_k | S_k)}$, called importance sampling ratio,
and the key matrix $\textbf{A}_{\text{off}}$ of off-policy TD(0) is % All components of the matrix $\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are positive.
\begin{equation*} % The row sums of $\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are positive. And The row sums of
\textbf{A}_{\text{off}} = \bm{\Phi}^{\top}\textbf{D}_{\mu}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}. % $\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are
\end{equation*} % \begin{equation*}
where $\textbf{D}_{\mu}$ is the $N \times N$ diagonal % \begin{array}{ccl}
matrix with $\textbf{d}_{\mu}$ on its diagonal. $\textbf{d}_{\mu}$ is a vector, each component representing the steady-state % \textbf{1}^{\top}\textbf{D}_{\pi}(\textbf{I}-\gamma \textbf{P}_{\pi})&=&\textbf{d}_{\pi}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
distribution under $\mu$ % &=& \textbf{d}_{\pi}^{\top} - \gamma \textbf{d}_{\pi}^{\top}\textbf{P}_{\pi}\\
% &=& \textbf{d}_{\pi}^{\top} - \gamma \textbf{d}_{\pi}^{\top}\\
If the key matrix % &=& (1-\gamma)\textbf{d}_{\pi}^{\top},
$\textbf{A}$ in the algorithm is positive definite, then the % \end{array}
algorithm is stable and converges. However, in the off-policy TD(0) % \end{equation*}
algorithm, it cannot be guaranteed that % all components of which are positive. Thus, the key matrix and its $\textbf{A}_{\text{on}}$ matrix are positive
$\textbf{A}$ is a positive definite matrix. In the 2-state counterexample, % definite, and on-policy TD(0) is stable
$\textbf{A}_{\text{off}}=-0.2$, which means that off-policy TD(0) cannot stably converge.
% The off-policy TD(0) update formula is
TDC and ETD are two well-known off-policy algorithms. % \begin{equation*}
The former is an off-policy algorithm derived from the % \label{thetatd_offpolicy}
objective function Mean Squared Projected Bellman error (MSPBE), while the latter employs a technique % \begin{array}{ccl}
to transform the key matrix % \bm{\theta}_{k+1}&\leftarrow&\bm{\theta}_k+\alpha_k \rho_k \delta_k\bm{\phi}_k,
$\textbf{A}$ in the original off-policy TD(0) from non-positive % \end{array}
definite to positive definite, thereby ensuring the algorithm's % \end{equation*}
convergence under off-policy conditions. % where $\rho_k =\frac{\pi(A_k | S_k)}{\mu(A_k | S_k)}$, called importance sampling ratio, and
The key matrix $\textbf{A}_{\text{off}}$ of off-policy TD(0) is
The MSPBE with importance sampling is
\begin{equation*}
\begin{array}{ccl}
\text{MSPBE}(\bm{\theta})&=&||\textbf{V}_{\bm{\theta}} - \Pi \textbf{T}^{\pi}\textbf{V}_{\bm{\theta}}||^{2}_{\mu}\\
&=&||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}^{\pi}\textbf{V}_{\bm{\theta}})||^{2}_{\mu}\\
&=&\mathbb{E}[\rho \delta \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[\rho \delta \bm{\phi}],
\end{array}
\end{equation*}
where $\textbf{V}_{\bm{\theta}}$ is viewed as vectors with one element for each state,
the norm $||\bm{v}||^{2}_{\mu}=\sum_{s}^{}\mu(s)\bm{v}^{2}(s)$, $\textbf{T}^{\pi}$, simplified to
$\textbf{T}$ in the following text, is Bellman operator and $\bm{\Pi}=\bm{\Phi}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1}\bm{\Phi}^{\top}\textbf{D}$.
The TDC update formula with importance sampling is
\begin{equation*}
\bm{\theta}_{k+1}\leftarrow\bm{\theta}_{k}+\alpha_{k} \rho_{k}[\delta_{k} \bm{\phi}_k- \gamma\bm{\phi}_{k+1}(\bm{\phi}^{\top}_k \bm{u}_{k})],
\label{thetatdc}
\end{equation*}
\begin{equation*} \begin{equation*}
\bm{u}_{k+1}\leftarrow \bm{u}_{k}+\zeta_{k}[\rho_k \delta_{k} - \bm{\phi}^{\top}_k \bm{u}_{k}]\bm{\phi}_k. \textbf{A}_{\text{off}} = \bm{\Phi}^{\top}\textbf{D}_{\mu}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi},
\label{utdc}
\end{equation*} \end{equation*}
where $\textbf{D}_{\mu}$ is the $|S| \times |S|$ diagonal
matrix with $d_{\mu}$ on its diagonal. $d_{\mu}$ is a vector, each component representing the steady-state
distribution under behavior policy $\mu$.
% If the key matrix
% $\textbf{A}$ in the algorithm is positive definite, then the
% algorithm is stable and converges. However, in the off-policy TD(0)
% algorithm, it cannot be guaranteed that
% $\textbf{A}$ is a positive definite matrix.
In the off-policy 2-state,
$\textbf{A}_{\text{off}}=-0.2$, which means that off-policy TD(0) cannot stably converge,
while , in the on-policy 2-state, $\textbf{A}_{\text{on}}=0.475$, which means that on-policy TD(0) can stably converge.
% TDC and ETD are two well-known off-policy algorithms.
% The former is an off-policy algorithm derived from the
% objective function Mean Squared Projected Bellman error (MSPBE), while the latter employs a technique
% to transform the key matrix
% $\textbf{A}$ in the original off-policy TD(0) from non-positive
% definite to positive definite, thereby ensuring the algorithm's
% convergence under off-policy conditions.
% The MSPBE with importance sampling is
% \begin{equation*}
% \begin{array}{ccl}
% \text{MSPBE}(\bm{\theta})&=&||\textbf{V}_{\bm{\theta}} - \Pi \textbf{T}^{\pi}\textbf{V}_{\bm{\theta}}||^{2}_{\mu}\\
% &=&||\Pi(\textbf{V}_{\bm{\theta}} - \textbf{T}^{\pi}\textbf{V}_{\bm{\theta}})||^{2}_{\mu}\\
% &=&\mathbb{E}[\rho \delta \bm{\phi}]^{\top} \mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[\rho \delta \bm{\phi}],
% \end{array}
% \end{equation*}
% where $\textbf{V}_{\bm{\theta}}$ is viewed as vectors with one element for each state,
% the norm $||\bm{v}||^{2}_{\mu}=\sum_{s}^{}\mu(s)\bm{v}^{2}(s)$, $\textbf{T}^{\pi}$, simplified to
% $\textbf{T}$ in the following text, is Bellman operator and $\bm{\Pi}=\bm{\Phi}(\bm{\Phi}^{\top}\textbf{D}\bm{\Phi})^{-1}\bm{\Phi}^{\top}\textbf{D}$.
% The TDC update formula with importance sampling is
% \begin{equation*}
% \bm{\theta}_{k+1}\leftarrow\bm{\theta}_{k}+\alpha_{k} \rho_{k}[\delta_{k} \bm{\phi}_k- \gamma\bm{\phi}_{k+1}(\bm{\phi}^{\top}_k \bm{u}_{k})],
% \label{thetatdc}
% \end{equation*}
% \begin{equation*}
% \bm{u}_{k+1}\leftarrow \bm{u}_{k}+\zeta_{k}[\rho_k \delta_{k} - \bm{\phi}^{\top}_k \bm{u}_{k}]\bm{\phi}_k.
% \label{utdc}
% \end{equation*}
The key matrix $\textbf{A}_{\text{TDC}}= \textbf{A}^{\top}_{\text{off}}\textbf{C}^{-1}\textbf{A}_{\text{off}}$, The key matrix $\textbf{A}_{\text{TDC}}= \textbf{A}^{\top}_{\text{off}}\textbf{C}^{-1}\textbf{A}_{\text{off}}$,
where $\textbf{C}=\mathbb{E}[\bm{\bm{\phi}}\bm{\bm{\phi}}^{\top}]$. where $\textbf{C}=\mathbb{E}[\bm{\bm{\phi}}\bm{\bm{\phi}}^{\top}]$.
In the 2-state counterexample, In the 2-state counterexample,
$\textbf{A}_{\text{TDC}}=0.016$, which means that TDC can stably converge. $\textbf{A}_{\text{TDC}}=0.016$, which means that TDC can stably converge.
The ETD update formula is The key matrix $\textbf{A}_{\text{TDC}}$ of on-policy TDC is
\begin{equation}
\label{fvmetd}
F_k \leftarrow \gamma \rho_{k-1}F_{k-1}+1,
\end{equation}
\begin{equation*} \begin{equation*}
\label{thetaetd} \textbf{A}_{\text{TDC}} = \textbf{A}^{\top}_{\text{on}}\textbf{C}^{-1}\textbf{A}_{\text{on}}.
\bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\alpha_k F_k \rho_k\delta_k\bm{\phi}_k, \end{equation*}
The key matrix $\textbf{A}_{\text{TDC}}$ of off-policy TDC is
\begin{equation*}
\textbf{A}_{\text{TDC}} = \textbf{A}^{\top}_{\text{off}}\textbf{C}^{-1}\textbf{A}_{\text{off}}.
\end{equation*}
$\textbf{A}_{\text{TDC}}=0.016$ in the off-policy 2-state and $\textbf{A}_{\text{TDC}}=0.09025$
in the on-policy 2-state, which means that TDC can stably converge in two settings.
To address the issue of the key matrix $\textbf{A}_{\text{off}}$
in off-policy TD(0) being non-positive definite,
a scalar variable, $F_t$,
is introduced to obtain the off-policy TD(0) algorithm,
which ensures convergence under off-policy
conditions.
The key matrix $\textbf{A}_{\text{ETD}}$ is
\begin{equation*}
\textbf{A}_{\text{ETD}} = \bm{\Phi}^{\top}\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi},
\end{equation*} \end{equation*}
where $F_t$ is a scalar variable and $F_0=1$.
The key matrix $\textbf{A}_{\text{ETD}}= \bm{\Phi}^{\top}\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$,
where where
$\textbf{F}$ is a diagonal matrix with diagonal elements $\textbf{F}$ is a diagonal matrix with diagonal elements
$f(s)\dot{=}d_{\mu}(s)\lim_{t\rightarrow \infty}\mathbb{E}_{\mu}[F_k|S_k=s]$, $f(s)\dot{=}d_{\mu}(s)\lim_{t\rightarrow \infty}\mathbb{E}_{\mu}[F_t|S_t=s]$,
which we assume exists. which we assume exists.
The vector $\textbf{f}\in \mathbb{R}^N$ with components The vector $\textbf{f}\in \mathbb{R}^N$ with components
$[\textbf{f}]_s\dot{=}f(s)$ can be written as $[\textbf{f}]_s\dot{=}f(s)$ can be written as
...@@ -193,21 +264,71 @@ $[\textbf{f}]_s\dot{=}f(s)$ can be written as ...@@ -193,21 +264,71 @@ $[\textbf{f}]_s\dot{=}f(s)$ can be written as
\textbf{f}&=\textbf{d}_{\mu}+\gamma \textbf{P}_{\pi}^{\top}\textbf{d}_{\mu}+(\gamma \textbf{P}_{\pi}^{\top})^2\textbf{d}_{\mu}+\ldots\\ \textbf{f}&=\textbf{d}_{\mu}+\gamma \textbf{P}_{\pi}^{\top}\textbf{d}_{\mu}+(\gamma \textbf{P}_{\pi}^{\top})^2\textbf{d}_{\mu}+\ldots\\
&=(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}\textbf{d}_{\mu}. &=(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}\textbf{d}_{\mu}.
\end{split} \end{split}
\end{equation*}.
The row sums of
$\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are
\begin{equation*}
\begin{array}{ccl}
\textbf{1}^{\top}\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})&=&\textbf{f}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
&=& \textbf{d}_{\mu}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})^{-1}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
&=& \textbf{d}_{\mu}^{\top},
\end{array}
\end{equation*} \end{equation*}
and in the 2-state counterexample, In the off-policy 2-state,
$\textbf{A}_{\text{ETD}}=3.4$, which means that ETD can stably converge. $\textbf{A}_{\text{ETD}}=3.4$, which means that ETD can stably converge.
The convergence rate of the algorithm is related to the matrix
$\textbf{A}$. The larger the minimum eigenvalue of Table \ref{tab:min_eigenvalues} shows Minimum eigenvalues
$\textbf{A}$, the faster the convergence rate. In the 2-state case, the minimum eigenvalue of the matrix of various algorithms in the 2-state counterexample.
$\textbf{A}$ in ETD is the largest, so it converges the fastest.
Based on this theorem, can we derive an algorithm with a larger minimum eigenvalue for matrix $\textbf{A}$. In the on-policy 2-state environment, the minimum eigenvalue
of the key matrix for TDC is greater than that of TD(0),
indicating that TDC converges faster than TD(0) in this
environment. In the off-policy 2-state environment, the
minimum eigenvalue of the key matrix for ETD is the largest,
suggesting that ETD has the fastest convergence rate.
Minimum eigenvalue larger, algorithm's convergence faster.
To derive an algorithm with a larger minimum eigenvalue for matrix
$\textbf{A}$, it is necessary to propose new objective functions.
The mentioned objective functions in the Introduction
are all forms of error. Is minimizing error the only option
for value-based reinforcement learning?
Based on this observation,
we propose alternative objective functions instead of minimizing errors.
% The ETD update formula is
% \begin{equation}
% \label{fvmetd}
% F_k \leftarrow \gamma \rho_{k-1}F_{k-1}+1,
% \end{equation}
% \begin{equation*}
% \label{thetaetd}
% \bm{\theta}_{k+1}\leftarrow \bm{\theta}_k+\alpha_k F_k \rho_k\delta_k\bm{\phi}_k,
% \end{equation*}
% where $F_t$ is a scalar variable and $F_0=1$.
% The key matrix $\textbf{A}_{\text{ETD}}= \bm{\Phi}^{\top}\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})\bm{\Phi}$,
% where
% $\textbf{F}$ is a diagonal matrix with diagonal elements
% $f(s)\dot{=}d_{\mu}(s)\lim_{t\rightarrow \infty}\mathbb{E}_{\mu}[F_k|S_k=s]$,
% which we assume exists.
% The vector $\textbf{f}\in \mathbb{R}^N$ with components
% $[\textbf{f}]_s\dot{=}f(s)$ can be written as
% \begin{equation*}
% \begin{split}
% \textbf{f}&=\textbf{d}_{\mu}+\gamma \textbf{P}_{\pi}^{\top}\textbf{d}_{\mu}+(\gamma \textbf{P}_{\pi}^{\top})^2\textbf{d}_{\mu}+\ldots\\
% &=(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}\textbf{d}_{\mu}.
% \end{split}
% \end{equation*}.
% The row sums of
% $\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})$ are
% \begin{equation*}
% \begin{array}{ccl}
% \textbf{1}^{\top}\textbf{F}(\textbf{I}-\gamma \textbf{P}_{\pi})&=&\textbf{f}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
% &=& \textbf{d}_{\mu}^{\top}(\textbf{I}-\gamma \textbf{P}_{\pi})^{-1}(\textbf{I}-\gamma \textbf{P}_{\pi})\\
% &=& \textbf{d}_{\mu}^{\top},
% \end{array}
% \end{equation*}
% and in the 2-state counterexample,
% $\textbf{A}_{\text{ETD}}=3.4$, which means that ETD can stably converge.
% In the 2-state case, the minimum eigenvalue of the matrix
% $\textbf{A}$ in ETD is the largest, so it converges the fastest.
% Based on this theorem, can we derive an algorithm with a larger minimum eigenvalue for matrix $\textbf{A}$.
NEW_aaai/main/theory.tex
\section{Theoretical Analysis} \section{Theoretical Analysis}
The purpose of this section is to establish the stabilities of the VMTDC algorithm This section primarily focuses on proving the convergence of VMTD, VMTDC, and VMETD.
and the VMETD algorithm. \begin{theorem}
\label{theorem1}(Convergence of VMTD).
In the case of on-policy learning, consider the iterations (\ref{omega}) and (\ref{theta}) with (\ref{delta}) of VMTD.
Let the step-size sequences $\alpha_k$ and $\beta_k$, $k\geq 0$ satisfy in this case $\alpha_k,\beta_k>0$, for all $k$,
$
\sum_{k=0}^{\infty}\alpha_k=\sum_{k=0}^{\infty}\beta_k=\infty,
$
$
\sum_{k=0}^{\infty}\alpha_k^2<\infty,
$
$
\sum_{k=0}^{\infty}\beta_k^2<\infty,
$
and
$
\alpha_k = o(\beta_k).
$
Assume that $(\phi_k,r_k,\phi_k')$ is an i.i.d. sequence with
uniformly bounded second moments, where $\phi_k$ and $\phi'_{k}$ are sampled from the same Markov chain.
Let $\textbf{A} = \mathrm{Cov}(\phi,\phi-\gamma\phi')$,
$b=\mathrm{Cov}(r,\phi)$.
Assume that matrix $\textbf{A}$ is non-singular.
Then the parameter vector $\theta_k$ converges with probability one
to $\textbf{A}^{-1}b$.
\end{theorem}
\begin{proof}
\label{th1proof}
The proof is based on Borkar's Theorem for
general stochastic approximation recursions with two time scales
\cite{borkar1997stochastic}.
A sketch proof is given as follows.
In the fast time scale, the parameter $w$ converges to
$\mathbb{E}[\delta|\theta_k]$.
In the slow time scale,
the associated ODE is
\begin{equation*}
\vec{h}(\theta(t))=-\textbf{A}\theta(t)+b.
\end{equation*}
\begin{equation}
\begin{array}{ccl}
A &=& \mathrm{Cov}(\phi,\phi-\gamma\phi')\\
&=&\frac{\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')-\mathrm{Cov}(\gamma\phi',\gamma\phi')}{2}\\
&=&\frac{\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')-\gamma^2\mathrm{Cov}(\phi',\phi')}{2}\\
&=&\frac{(1-\gamma^2)\mathrm{Cov}(\phi,\phi)+\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')}{2},\\
\end{array}
\label{covariance}
\end{equation}
where we eventually used $\mathrm{Cov}(\phi',\phi')=\mathrm{Cov}(\phi,\phi)$
\footnote{The covariance matrix $\mathrm{Cov}(\phi',\phi')$ is equal to
the covariance matrix $\mathrm{Cov}(\phi,\phi)$ if the initial state is re-reachable or
initialized randomly in a Markov chain for on-policy update.}.
Note that the covariance matrix $\mathrm{Cov}(\phi,\phi)$ and
$\mathrm{Cov}(\phi-\gamma\phi',\phi-\gamma\phi')$ are semi-positive
definite. Then, the matrix $\textbf{A}$ is semi-positive definite because $\textbf{A}$ is
linearly combined by two positive-weighted semi-positive definite matrice
(\ref{covariance}).
Furthermore, $\textbf{A}$ is nonsingular due to the assumption.
Hence, the matrix $\textbf{A}$ is positive definite. And,
the parameter $\theta$ converges to $\textbf{A}^{-1}b$.
\end{proof}
Please refer to the appendix for VMTD's detailed proof process.
\begin{theorem} \begin{theorem}
\label{theorem2}(Convergence of VMTDC). \label{theorem2}(Convergence of VMTDC).
In the case of off-policy learning, consider the iterations (\ref{omegavmtdc}), (\ref{uvmtdc}) and (\ref{thetavmtdc}) of VMTDC. In the case of off-policy learning, consider the iterations (\ref{omegavmtdc}), (\ref{uvmtdc}) and (\ref{thetavmtdc}) of VMTDC.
...@@ -24,129 +87,147 @@ and the VMETD algorithm. ...@@ -24,129 +87,147 @@ and the VMETD algorithm.
$ $
\zeta_k = o(\beta_k). \zeta_k = o(\beta_k).
$ $
Assume that $(\bm{\bm{\phi}}_k,r_k,\bm{\bm{\phi}}_k')$ is an i.i.d. sequence with Assume that $(\phi_k,r_k,\phi_k')$ is an i.i.d. sequence with
uniformly bounded second moments. uniformly bounded second moments.
Let $\textbf{A} = \mathrm{Cov}(\bm{\bm{\phi}},\bm{\bm{\phi}}-\gamma\bm{\bm{\phi}}')$, Let $\textbf{A} = \mathrm{Cov}(\phi,\phi-\gamma\phi')$,
$\bm{b}=\mathrm{Cov}(r,\bm{\bm{\phi}})$, and $\textbf{C}=\mathbb{E}[\bm{\bm{\phi}}\bm{\bm{\phi}}^{\top}]$. $b=\mathrm{Cov}(r,\phi)$, and $\textbf{C}=\mathbb{E}[\phi\phi^{\top}]$.
Assume that $\textbf{A}$ and $\textbf{C}$ are non-singular matrices. Assume that $\textbf{A}$ and $\textbf{C}$ are non-singular matrices.
Then the parameter vector $\bm{\theta}_k$ converges with probability one Then the parameter vector $\theta_k$ converges with probability one
to $\textbf{A}^{-1}\bm{b}$. to $\textbf{A}^{-1}b$.
\end{theorem} \end{theorem}
\begin{proof} \begin{proof}
The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation. The proof is similar to that given by \cite{sutton2009fast} for TDC,
% For the VMTDC algorithm, a new one-step linear TD solution is defined as: but it is based on multi-time-scale stochastic approximation.
% \begin{equation*}
% 0=\mathbb{E}[(\bm{\phi} - \gamma \bm{\phi}' - \mathbb{E}[\bm{\phi} - \gamma \bm{\phi}'])\bm{\phi}^\top]\mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta])\bm{\phi}]=\textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}\bm{\theta}+\bm{b}). A sketch proof is given as follows.
% \end{equation*} In the fastest time scale, the parameter $w$ converges to
% The matrix $\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}$ is positive definite. Thus, the VMTD's solution is $\mathbb{E}[\delta|u_k,\theta_k]$.
% $\bm{\theta}_{\text{VMTDC}}=\bm{\theta}_{\text{VMTD}}=\textbf{A}^{-1}\bm{b}$. In the second fast time scale,
the parameter $u$ converges to $\textbf{C}^{-1}\mathbb{E}[(\delta-\mathbb{E}[\delta|\theta_k])\phi|\theta_k]$.
First, note that recursion (\ref{thetavmtdc}) and (\ref{uvmtdc}) can be rewritten as, respectively, In the slower time scale,
\begin{equation*} the associated ODE is
\theta_{k+1}\leftarrow \theta_k+\zeta_k x(k), \begin{equation*}
\end{equation*} \vec{h}(\theta(t))=\textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}\theta(t)+b).
\begin{equation*} \end{equation*}
u_{k+1}\leftarrow u_k+\beta_k y(k), The matrix $\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}$ is positive definite. Thus,
\end{equation*} the parameter $\theta$ converges to $\textbf{A}^{-1}b$.
where \end{proof}
\begin{equation*} Please refer to the appendix for VMTDC's detailed proof process.
x(k)=\frac{\alpha_k}{\zeta_k}[(\delta_{k}- \omega_k) \phi_k - \gamma\phi'_{k}(\phi^{\top}_k u_k)], % \begin{proof}
\end{equation*} % The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation.
\begin{equation*} % % For the VMTDC algorithm, a new one-step linear TD solution is defined as:
y(k)=\frac{\zeta_k}{\beta_k}[\delta_{k}-\omega_k - \phi^{\top}_k u_k]\phi_k. % % \begin{equation*}
\end{equation*} % % 0=\mathbb{E}[(\bm{\phi} - \gamma \bm{\phi}' - \mathbb{E}[\bm{\phi} - \gamma \bm{\phi}'])\bm{\phi}^\top]\mathbb{E}[\bm{\phi} \bm{\phi}^{\top}]^{-1}\mathbb{E}[(\delta -\mathbb{E}[\delta])\bm{\phi}]=\textbf{A}^{\top}\textbf{C}^{-1}(-\textbf{A}\bm{\theta}+\bm{b}).
% % \end{equation*}
% % The matrix $\textbf{A}^{\top}\textbf{C}^{-1}\textbf{A}$ is positive definite. Thus, the VMTD's solution is
% % $\bm{\theta}_{\text{VMTDC}}=\bm{\theta}_{\text{VMTD}}=\textbf{A}^{-1}\bm{b}$.
Recursion (\ref{thetavmtdc}) can also be rewritten as % First, note that recursion (\ref{thetavmtdc}) and (\ref{uvmtdc}) can be rewritten as, respectively,
\begin{equation*} % \begin{equation*}
\theta_{k+1}\leftarrow \theta_k+\beta_k z(k), % \theta_{k+1}\leftarrow \theta_k+\zeta_k x(k),
\end{equation*} % \end{equation*}
where % \begin{equation*}
\begin{equation*} % u_{k+1}\leftarrow u_k+\beta_k y(k),
z(k)=\frac{\alpha_k}{\beta_k}[(\delta_{k}- \omega_k) \phi_k - \gamma\phi'_{k}(\phi^{\top}_k u_k)], % \end{equation*}
\end{equation*} % where
% \begin{equation*}
% x(k)=\frac{\alpha_k}{\zeta_k}[(\delta_{k}- \omega_k) \phi_k - \gamma\phi'_{k}(\phi^{\top}_k u_k)],
% \end{equation*}
% \begin{equation*}
% y(k)=\frac{\zeta_k}{\beta_k}[\delta_{k}-\omega_k - \phi^{\top}_k u_k]\phi_k.
% \end{equation*}
Due to the settings of step-size schedule % Recursion (\ref{thetavmtdc}) can also be rewritten as
$\alpha_k = o(\zeta_k)$, $\zeta_k = o(\beta_k)$, $x(k)\rightarrow 0$, $y(k)\rightarrow 0$, $z(k)\rightarrow 0$ almost surely as $k\rightarrow 0$. % \begin{equation*}
That is that the increments in iteration (\ref{omegavmtdc}) are uniformly larger than % \theta_{k+1}\leftarrow \theta_k+\beta_k z(k),
those in (\ref{uvmtdc}) and the increments in iteration (\ref{uvmtdc}) are uniformly larger than % \end{equation*}
those in (\ref{thetavmtdc}), thus (\ref{omegavmtdc}) is the fastest recursion, (\ref{uvmtdc}) is the second fast recursion and (\ref{thetavmtdc}) is the slower recursion. % where
Along the fastest time scale, iterations of (\ref{thetavmtdc}), (\ref{uvmtdc}) and (\ref{omegavmtdc}) % \begin{equation*}
are associated to ODEs system as follows: % z(k)=\frac{\alpha_k}{\beta_k}[(\delta_{k}- \omega_k) \phi_k - \gamma\phi'_{k}(\phi^{\top}_k u_k)],
\begin{equation} % \end{equation*}
\dot{\theta}(t) = 0,
\label{thetavmtdcFastest}
\end{equation}
\begin{equation}
\dot{u}(t) = 0,
\label{uvmtdcFastest}
\end{equation}
\begin{equation}
\dot{\omega}(t)=\mathbb{E}[\delta_t|u(t),\theta(t)]-\omega(t).
\label{omegavmtdcFastest}
\end{equation}
Based on the ODE (\ref{thetavmtdcFastest}) and (\ref{uvmtdcFastest}), both $\theta(t)\equiv \theta$ % Due to the settings of step-size schedule
and $u(t)\equiv u$ when viewed from the fastest timescale. % $\alpha_k = o(\zeta_k)$, $\zeta_k = o(\beta_k)$, $x(k)\rightarrow 0$, $y(k)\rightarrow 0$, $z(k)\rightarrow 0$ almost surely as $k\rightarrow 0$.
By the Hirsch lemma \cite{hirsch1989convergent}, it follows that % That is that the increments in iteration (\ref{omegavmtdc}) are uniformly larger than
$||\theta_k-\theta||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some % those in (\ref{uvmtdc}) and the increments in iteration (\ref{uvmtdc}) are uniformly larger than
$\theta$ that depends on the initial condition $\theta_0$ of recursion % those in (\ref{thetavmtdc}), thus (\ref{omegavmtdc}) is the fastest recursion, (\ref{uvmtdc}) is the second fast recursion and (\ref{thetavmtdc}) is the slower recursion.
(\ref{thetavmtdc}) and $||u_k-u||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some % Along the fastest time scale, iterations of (\ref{thetavmtdc}), (\ref{uvmtdc}) and (\ref{omegavmtdc})
$u$ that depends on the initial condition $u_0$ of recursion % are associated to ODEs system as follows:
(\ref{uvmtdc}). Thus, the ODE pair (\ref{thetavmtdcFastest})-(ref{omegavmtdcFastest}) % \begin{equation}
can be written as % \dot{\theta}(t) = 0,
\begin{equation} % \label{thetavmtdcFastest}
\dot{\omega}(t)=\mathbb{E}[\delta_t|u,\theta]-\omega(t). % \end{equation}
\label{omegavmtdcFastestFinal} % \begin{equation}
\end{equation} % \dot{u}(t) = 0,
% \label{uvmtdcFastest}
% \end{equation}
% \begin{equation}
% \dot{\omega}(t)=\mathbb{E}[\delta_t|u(t),\theta(t)]-\omega(t).
% \label{omegavmtdcFastest}
% \end{equation}
Consider the function $h(\omega)=\mathbb{E}[\delta|\theta,u]-\omega$, % Based on the ODE (\ref{thetavmtdcFastest}) and (\ref{uvmtdcFastest}), both $\theta(t)\equiv \theta$
i.e., the driving vector field of the ODE (\ref{omegavmtdcFastestFinal}). % and $u(t)\equiv u$ when viewed from the fastest timescale.
It is easy to find that the function $h$ is Lipschitz with coefficient % By the Hirsch lemma \cite{hirsch1989convergent}, it follows that
$-1$. % $||\theta_k-\theta||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
Let $h_{\infty}(\cdot)$ be the function defined by % $\theta$ that depends on the initial condition $\theta_0$ of recursion
$h_{\infty}(\omega)=\lim_{r\rightarrow \infty}\frac{h(r\omega)}{r}$. % (\ref{thetavmtdc}) and $||u_k-u||\rightarrow 0$ a.s. as $k\rightarrow \infty$ for some
Then $h_{\infty}(\omega)= -\omega$, is well-defined. % $u$ that depends on the initial condition $u_0$ of recursion
For (\ref{omegavmtdcFastestFinal}), $\omega^*=\mathbb{E}[\delta|\theta,u]$ % (\ref{uvmtdc}). Thus, the ODE pair (\ref{thetavmtdcFastest})-(ref{omegavmtdcFastest})
is the unique globally asymptotically stable equilibrium. % can be written as
For the ODE % \begin{equation}
\begin{equation} % \dot{\omega}(t)=\mathbb{E}[\delta_t|u,\theta]-\omega(t).
\dot{\omega}(t) = h_{\infty}(\omega(t))= -\omega(t), % \label{omegavmtdcFastestFinal}
\label{omegavmtdcInfty} % \end{equation}
\end{equation}
apply $\vec{V}(\omega)=(-\omega)^{\top}(-\omega)/2$ as its % Consider the function $h(\omega)=\mathbb{E}[\delta|\theta,u]-\omega$,
associated strict Liapunov function. Then, % i.e., the driving vector field of the ODE (\ref{omegavmtdcFastestFinal}).
the origin of (\ref{omegavmtdcInfty}) is a globally asymptotically stable % It is easy to find that the function $h$ is Lipschitz with coefficient
equilibrium. % $-1$.
% Let $h_{\infty}(\cdot)$ be the function defined by
Consider now the recursion (\ref{omegavmtdc}). % $h_{\infty}(\omega)=\lim_{r\rightarrow \infty}\frac{h(r\omega)}{r}$.
Let % Then $h_{\infty}(\omega)= -\omega$, is well-defined.
$M_{k+1}=(\delta_k-\omega_k) % For (\ref{omegavmtdcFastestFinal}), $\omega^*=\mathbb{E}[\delta|\theta,u]$
-\mathbb{E}[(\delta_k-\omega_k)|\mathcal{F}(k)]$, % is the unique globally asymptotically stable equilibrium.
where $\mathcal{F}(k)=\sigma(\omega_l,u_l,\theta_l,l\leq k;\phi_s,\phi_s',r_s,s<k)$, % For the ODE
$k\geq 1$ are the sigma fields % \begin{equation}
generated by $\omega_0,u_0,\theta_0,\omega_{l+1},u_{l+1},\theta_{l+1},\phi_l,\phi_l'$, % \dot{\omega}(t) = h_{\infty}(\omega(t))= -\omega(t),
$0\leq l<k$. % \label{omegavmtdcInfty}
It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that % \end{equation}
satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$. % apply $\vec{V}(\omega)=(-\omega)^{\top}(-\omega)/2$ as its
Because $\phi_k$, $r_k$, and $\phi_k'$ have % associated strict Liapunov function. Then,
uniformly bounded second moments, it can be seen that for some constant % the origin of (\ref{omegavmtdcInfty}) is a globally asymptotically stable
$c_1>0$, $\forall k\geq0$, % equilibrium.
\begin{equation*}
\mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq
c_1(1+||\omega_k||^2+||u_k||^2+||\theta_k||^2).
\end{equation*}
% Consider now the recursion (\ref{omegavmtdc}).
% Let
% $M_{k+1}=(\delta_k-\omega_k)
% -\mathbb{E}[(\delta_k-\omega_k)|\mathcal{F}(k)]$,
% where $\mathcal{F}(k)=\sigma(\omega_l,u_l,\theta_l,l\leq k;\phi_s,\phi_s',r_s,s<k)$,
% $k\geq 1$ are the sigma fields
% generated by $\omega_0,u_0,\theta_0,\omega_{l+1},u_{l+1},\theta_{l+1},\phi_l,\phi_l'$,
% $0\leq l<k$.
% It is easy to verify that $M_{k+1},k\geq0$ are integrable random variables that
% satisfy $\mathbb{E}[M_{k+1}|\mathcal{F}(k)]=0$, $\forall k\geq0$.
% Because $\phi_k$, $r_k$, and $\phi_k'$ have
% uniformly bounded second moments, it can be seen that for some constant
% $c_1>0$, $\forall k\geq0$,
% \begin{equation*}
% \mathbb{E}[||M_{k+1}||^2|\mathcal{F}(k)]\leq
% c_1(1+||\omega_k||^2+||u_k||^2+||\theta_k||^2).
% \end{equation*}
Now Assumptions (A1) and (A2) of \cite{borkar2000ode} are verified.
Furthermore, Assumptions (TS) of \cite{borkar2000ode} is satisfied by our
conditions on the step-size sequences $\alpha_k$,$\zeta_k$, $\beta_k$. Thus,
by Theorem 2.2 of \cite{borkar2000ode} we obtain that
$||\omega_k-\omega^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$.
Recursion (\ref{uvmtdc}) is considered the second timescale. % Now Assumptions (A1) and (A2) of \cite{borkar2000ode} are verified.
Recursion (\ref{thetavmtdc}) is considered the slower timescale. % Furthermore, Assumptions (TS) of \cite{borkar2000ode} is satisfied by our
For the convergence properties of $u$ and $\theta$, please refer to the appendix. % conditions on the step-size sequences $\alpha_k$,$\zeta_k$, $\beta_k$. Thus,
\end{proof} % by Theorem 2.2 of \cite{borkar2000ode} we obtain that
% $||\omega_k-\omega^*||\rightarrow 0$ almost surely as $k\rightarrow \infty$.
% Recursion (\ref{uvmtdc}) is considered the second timescale.
% Recursion (\ref{thetavmtdc}) is considered the slower timescale.
% For the convergence properties of $u$ and $\theta$, please refer to the appendix.
% \end{proof}
% \begin{proof} % \begin{proof}
% The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation. % The proof is similar to that given by \cite{sutton2009fast} for TDC, but it is based on multi-time-scale stochastic approximation.
...@@ -426,18 +507,20 @@ For the convergence properties of $u$ and $\theta$, please refer to the appendix ...@@ -426,18 +507,20 @@ For the convergence properties of $u$ and $\theta$, please refer to the appendix
$ $
Assume that $(\bm{\bm{\phi}}_k,r_k,\bm{\bm{\phi}}_k')$ is an i.i.d. sequence with Assume that $(\bm{\bm{\phi}}_k,r_k,\bm{\bm{\phi}}_k')$ is an i.i.d. sequence with
uniformly bounded second moments, where $\bm{\bm{\phi}}_k$ and $\bm{\bm{\phi}}'_{k}$ are sampled from the same Markov chain. uniformly bounded second moments, where $\bm{\bm{\phi}}_k$ and $\bm{\bm{\phi}}'_{k}$ are sampled from the same Markov chain.
Let $\textbf{A}_{\textbf{VMETD}} ={\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ){\bm{\Phi}}$, Let $\textbf{A}_{\textbf{VMETD}} ={\bm{\Phi}}^{\top} (\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-d_{\mu} d_{\mu}^{\top} ){\bm{\Phi}}$,
$\bm{b}_{\textbf{VMETD}}=\bm{\Phi}^{\top}(\textbf{F}-\textbf{d}_{\mu} \textbf{f}^{\top})\textbf{r}_{\pi}$. $b_{\text{VMETD}}=\bm{\Phi}^{\top}(\textbf{F}-d_{\mu} f^{\top})r_{\pi}$.
Assume that matrix $A$ is non-singular. Assume that matrix $\textbf{A}$ is non-singular.
Then the parameter vector $\bm{\theta}_k$ converges with probability one Then the parameter vector $\theta_k$ converges with probability one
to $\textbf{A}_{\textbf{VMETD}}^{-1}\bm{b}_{\textbf{VMETD}}$. to $\textbf{A}_{\textbf{VMETD}}^{-1}b_{\textbf{VMETD}}$.
\end{theorem} \end{theorem}
\begin{proof} \begin{proof}
The proof of VMETD's convergence is also based on Borkar's Theorem for The proof of VMETD's convergence is also based on Borkar's Theorem for
general stochastic approximation recursions with two time scales general stochastic approximation recursions with two time scales
\cite{borkar1997stochastic}. \cite{borkar1997stochastic}.
Recursion (\ref{omegavmetd}) is considered the faster timescale. For the convergence properties of $\omega$, please refer to the appendix. A sketch proof is given as follows.
In the fast time scale, the parameter $\omega$ converges to
$\mathbb{E}_{\mu}[F\rho\delta|\theta_k]$.
Recursion (\ref{thetavmetd}) is considered the slower timescale. Recursion (\ref{thetavmetd}) is considered the slower timescale.
If the key matrix If the key matrix
$\textbf{A}_{\text{VMETD}}$ is positive definite, then $\textbf{A}_{\text{VMETD}}$ is positive definite, then
...@@ -445,70 +528,37 @@ $\theta$ converges. ...@@ -445,70 +528,37 @@ $\theta$ converges.
\begin{equation} \begin{equation}
\label{rowsum} \label{rowsum}
\begin{split} \begin{split}
(\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} )\textbf{1} &(\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-{d}_{\mu} {d}_{\mu}^{\top} )e\\
&=\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})\textbf{1}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbf{1}\\ &=\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})e-{d}_{\mu} {d}_{\mu}^{\top} e\\
&=\textbf{F}(\textbf{1}-\gamma \textbf{P}_{\pi} \textbf{1})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbf{1}\\ % &=\textbf{F}(\textbfe-\gamma \textbf{P}_{\pi} \textbfe)-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbfe\\
&=(1-\gamma)\textbf{F}\textbf{1}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbf{1}\\ &=(1-\gamma)\textbf{F}e-{d}_{\mu} {d}_{\mu}^{\top} e\\
&=(1-\gamma)\textbf{f}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbf{1}\\ % &=(1-\gamma)\textbf{f}-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \textbfe\\
&=(1-\gamma)\textbf{f}-\textbf{d}_{\mu} \\ &=(1-\gamma){f}-{d}_{\mu} \\
&=(1-\gamma)(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}\textbf{d}_{\mu}-\textbf{d}_{\mu} \\ &=(1-\gamma)(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}{d}_{\mu}-{d}_{\mu} \\
&=(1-\gamma)[(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}-\textbf{I}]\textbf{d}_{\mu} \\ &=(1-\gamma)[(\textbf{I}-\gamma\textbf{P}_{\pi}^{\top})^{-1}-\textbf{I}]{d}_{\mu} \\
&=(1-\gamma)[\sum_{t=0}^{\infty}(\gamma\textbf{P}_{\pi}^{\top})^{t}-\textbf{I}]\textbf{d}_{\mu} \\ &=(1-\gamma)[\sum_{t=0}^{\infty}(\gamma\textbf{P}_{\pi}^{\top})^{t}-\textbf{I}]{d}_{\mu} \\
&=(1-\gamma)[\sum_{t=1}^{\infty}(\gamma\textbf{P}_{\pi}^{\top})^{t}]\textbf{d}_{\mu} > 0 \\ &=(1-\gamma)[\sum_{t=1}^{\infty}(\gamma\textbf{P}_{\pi}^{\top})^{t}]{d}_{\mu} > 0, \\
\end{split} \end{split}
\end{equation} \end{equation}
\begin{equation} \begin{equation}
\label{columnsum} \label{columnsum}
\begin{split} \begin{split}
\textbf{1}^{\top}(\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} ) &e^{\top}(\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} {d}_{\mu}^{\top} )\\
&=\textbf{1}^{\top}\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{1}^{\top}\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \\ &=e^{\top}\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-e^{\top}{d}_{\mu} {d}_{\mu}^{\top} \\
&=\textbf{d}_{\mu}^{\top}-\textbf{1}^{\top}\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top} \\ &={d}_{\mu}^{\top}-e^{\top}{d}_{\mu} {d}_{\mu}^{\top} \\
&=\textbf{d}_{\mu}^{\top}- \textbf{d}_{\mu}^{\top} \\ &={d}_{\mu}^{\top}- {d}_{\mu}^{\top} \\
&=0 &=0,
\end{split} \end{split}
\end{equation} \end{equation}
(\ref{rowsum}) and (\ref{columnsum}) show that the matrix $\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-\textbf{d}_{\mu} \textbf{d}_{\mu}^{\top}$ of where $e$ is the all-ones vector.
(\ref{rowsum}) and (\ref{columnsum}) show that the matrix $\textbf{F} (\textbf{I} - \gamma \textbf{P}_{\pi})-d_{\mu} d_{\mu}^{\top}$ of
diagonal entries are positive and its off-diagonal entries are negative. So its each row sum plus the corresponding column sum is positive. diagonal entries are positive and its off-diagonal entries are negative. So its each row sum plus the corresponding column sum is positive.
So $\textbf{A}_{\text{VMETD}}$ is positive definite. So $\textbf{A}_{\text{VMETD}}$ is positive definite.
\end{proof} \end{proof}
\begin{figure*}[htb]
\vskip 0.2in
\begin{center}
\subfigure[2-state counterexample]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/2-state.pdf}
\label{2-state}
}
\subfigure[7-state counterexample]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/7-state.pdf}
\label{7-state}
}
\subfigure[Maze]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/maze.pdf}
\label{MazeFull}
}\\
\subfigure[Cliff Walking]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/cl.pdf}
\label{CliffWalkingFull}
}
\subfigure[Mountain Car]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/mt.pdf}
\label{MountainCarFull}
}
\subfigure[Acrobot]{
\includegraphics[width=0.55\columnwidth, height=0.475\columnwidth]{main/pic/acrobot.pdf}
\label{AcrobotFull}
}
\caption{Learning curses of two evaluation environments and four contral environments.}
\label{Complete_full}
\end{center}
\vskip -0.2in
\end{figure*}
\subsection{Optimal Policy Invariance} \subsection{Optimal Policy Invariance}
This section prove This section prove
the optimal policy invariance of the optimal policy invariance of VMTD,
VMTDC and VMETD in control experiments, VMTDC and VMETD in control experiments,
laying the groundwork for subsequent experiments. laying the groundwork for subsequent experiments.
...@@ -518,7 +568,7 @@ true value and the predicted value, action $a_3$ is ...@@ -518,7 +568,7 @@ true value and the predicted value, action $a_3$ is
still chosen under the greedy-policy. still chosen under the greedy-policy.
On the contrary, supervised learning is usually used to predict temperature, humidity, morbidity, etc. If the bias is too large, the consequences could be serious. On the contrary, supervised learning is usually used to predict temperature, humidity, morbidity, etc. If the bias is too large, the consequences could be serious.
\begin{table}[t] \begin{table}[ht]
\caption{Comparison of action selection with and without \caption{Comparison of action selection with and without
constant bias in $Q$ values.} constant bias in $Q$ values.}
\label{example_bias} \label{example_bias}
......
画图.pptx 0 → 100644
File added
论文草稿.txt
题目：A Variance Minimization Approach to Off-policy Temporal-Difference Learning 题目：A Variance Minimization Approach to Off-policy Temporal-Difference Learning
...@@ -104,3 +104,29 @@ to, (1) 将标量参数引入到更多的TD算法中. ...@@ -104,3 +104,29 @@ to, (1) 将标量参数引入到更多的TD算法中.
对于控制实验，迷宫和cliff walking的实验结果相似，VMGQ表现优于GQ，EQ表现优于VMGQ，而VMEQ的性能最优。 对于控制实验，迷宫和cliff walking的实验结果相似，VMGQ表现优于GQ，EQ表现优于VMGQ，而VMEQ的性能最优。
mountain car和Acrobot的实验结果相似，VMGQ和VMEQ的性能接近都优于GQ和EQ。总之对于控制实验，VM算法优于非VM算法 mountain car和Acrobot的实验结果相似，VMGQ和VMEQ的性能接近都优于GQ和EQ。总之对于控制实验，VM算法优于非VM算法
接下来，我们将在2-state环境中计算TD(0)、TDC、ETD的分别在on-policy和off-policy下的各自A的最小特征值。
如果矩阵A正定，则算法收敛。
首先，我们将介绍2-state分别在on-policy和off-policy下的环境设定。
在on-policy设定下，行为策略与目标策略一样，令A=B。
为了解决off-policy TD(0)的关键矩阵A_off非正定问题，
为了方便
在2-state环境中，我们进行了两种实验——on-policy实验和off-policy实验，来验证算法的收敛速度与关键矩阵的最小特征值的关系。
图A是on-policy 2-state的策略评估实验的曲线图。在该实验设定下，TD、VMTD、TDC以及VMTDC的收敛速度在依次递减，而表1可以得到这四个算法的关键矩阵的最小特征值都大于0，并且依次递减。实验曲线和表格数值相照应。
图B是off-policy 2-state的策略评估实验的曲线图。在该实验设定下，ETD、VMETD、VMTD、VMTDC以及TDC的收敛速度在依次递减，TD则发散。而表1可以得到ETD、VMETD、VMTD、VMTDC以及TDC这五个算法的关键矩阵的最小特征值都大于0，并且依次递减，TD算法的关键矩阵的最小特征值小于0。实验曲线和表格数值相照应。令人惊喜的是，尽管VMTD是on-policy下保证收敛的算法，但在off-policy 2-state下依旧可以收敛。由VMTD的更新公式可以看出，VMTD的更新公式相当于是对TD更新的调整与修正，参数omega的引入使得梯度估计的方差更加稳定，从而让theta的更新更加稳定。
图1，2，3，4分别是四个控制实验的曲线图。四个控制实验都表现出了一个共性特征：VMEQ的表现优于EQ，VMGQ优于GQ，VMQ优于Q-learning，VMSarsa优于Sarsa。对于Maze和Cliffwalking实验，VMEQ都表现出了最佳的性能，收敛速度最快。对于Mountain car和 Acrobot实验，四个VM算法的表现近乎一样，并且都优于其他算法。
总的来说，不管是策略评估实验还是控制实验，VM算法都表现较为优秀，尤其在控制实验中特别突出。
在本论文中，
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment