This repository has been archived by the owner on Oct 4, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 3
/
lecture21.tex
233 lines (216 loc) · 8.4 KB
/
lecture21.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
%% LyX 2.2.1 created this file. For more info, see http://www.lyx.org/.
%% Do not edit unless you really know what you are doing.
\documentclass[english,12pt]{article}
\usepackage[T1]{fontenc}
\usepackage[latin9]{inputenc}
\usepackage{mathtools}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{amssymb}
\makeatletter
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Textclass specific LaTeX commands.
\theoremstyle{plain}
\newtheorem{thm}{\protect\theoremname}[section]
\theoremstyle{definition}
\newtheorem{defn}[thm]{\protect\definitionname}
\theoremstyle{plain}
\newtheorem{prop}[thm]{\protect\propositionname}
\ifx\proof\undefined
\newenvironment{proof}[1][\protect\proofname]{\par
\normalfont\topsep6\p@\@plus6\p@\relax
\trivlist
\itemindent\parindent
\item[\hskip\labelsep\scshape #1]\ignorespaces
}{%
\endtrivlist\@endpefalse
}
\providecommand{\proofname}{Proof}
\fi
\theoremstyle{remark}
\newtheorem{rem}[thm]{\protect\remarkname}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% User specified LaTeX commands.
\usepackage[margin=1in]{geometry}
\makeatother
\usepackage{babel}
\providecommand{\definitionname}{Definition}
\providecommand{\propositionname}{Proposition}
\providecommand{\remarkname}{Remark}
\providecommand{\theoremname}{Theorem}
\begin{document}
\title{Math 525: Lecture 21}
\date{April 03, 2018}
\maketitle
So far, we have worked primarily with (stationary) Markov chains whose
transition matrices are ``constant''. In this lecture, we explore
the following question: what if we could ``control'' the transition
matrix? In this context, we will have a transition matrix $P(\pi)$
that depends on some quantity $\pi$ which we, the ``controller'',
get to choose.
\section{Markov decision processes}
For this lecture, our setting is as follows:
\begin{itemize}
\item $S=\{1,\ldots,m\}$ is a finite state space.
\item To each state $i$ in $S$ is associated a nonempty countable set
$\mathcal{A}_{i}$ which we can intuitively think of as all the ``actions''
available at state $i$.
\end{itemize}
\begin{defn}
A \emph{stationary policy} $\pi_{0}$ is a function whose domain is
$S$ and which satisfies $\pi_{0}(i)\in\mathcal{A}_{i}$ for all $i$.
The set of all stationary policies is denoted $\Pi_{0}$
\end{defn}
%
\begin{defn}
A \emph{randomized policy} $(\pi_{n})_{n\geq0}$ is a sequence in
which each $\pi_{n}(i)$ is a random variable satisfying
\begin{itemize}
\item $\pi_{n}(i)$ takes values in $\mathcal{A}_{i}$ a.s. and
\item $\{\pi_{n}(i)=a\}\in\mathcal{F}_{n}$ for each $a$ in $\mathcal{A}_{i}$.
\end{itemize}
The set of all randomized policies is denoted $\Pi$.
\end{defn}
For each state $i$ in $S$ and action $a$ in $\mathcal{A}_{i}$,
let $p_{i}(a)$ denote a nonnegative column vector satisfying $p_{i}(a)^{\intercal}e=1$.
Given a randomized policy $\pi$, let $(X_{n}^{\pi})_{n\geq0}$ denote
a Markov chain satisfying
\[
\mathbb{P}(X_{n+1}^{\pi}=j\mid X_{n}^{\pi}=i)=p_{i}(\pi_{n}(i))^{\intercal}e_{j}.
\]
That is, the transition matrix at time $n$ is
\[
P(\pi_{n})=\begin{pmatrix}p_{1}(\pi_{n}(1))^{\intercal}\\
p_{2}(\pi_{n}(2))^{\intercal}\\
\vdots\\
p_{m}(\pi_{n}(m))^{\intercal}
\end{pmatrix}.
\]
Now, let $c:S\rightarrow\mathbb{R}$, $0\leq d<1$, and
\begin{equation}
J(i,\pi)=\mathbb{E}\left[\sum_{n\geq0}d^{n}c(X_{n}^{\pi})\middle|X_{0}^{\pi}=i\right].\label{eq:objective}
\end{equation}
We can think of
\begin{itemize}
\item $c(X_{n}^{\pi})$ as the cost incurred at time $n$ and
\item $d^{n}$ as a discount factor which attempts to capture the fact that
costs incurred in the ``future'' are not as bad as costs incurred
``today''.
\end{itemize}
Our objective is to pick $\pi$ so as to minimize $J(i,\pi)$. That
is, we are interested in the quantity
\begin{equation}
\boxed{v(i)=\inf_{\pi\in\Pi}J(i,\pi)}\label{eq:mdp}
\end{equation}
We call (\ref{eq:mdp}) a \emph{Markov decision process} (MDP).
\begin{prop}
$v(i)$ is bounded for each $i$.
\end{prop}
\begin{proof}
This is a trivial consequence of the discount factor being strictly
less than one:
\[
\left|v(i)\right|\leq\sum_{n\geq0}d^{n}\max_{j}\left|c(j)\right|=\frac{1}{1-d}\max_{j}\left|c(j)\right|.\qedhere
\]
\end{proof}
\begin{rem}
We glossed over defining $\mathcal{F}_{n}$ earlier, so we return
to that now. Given a particular randomized policy $\pi$, we define
\[
\mathcal{F}_{n}=\sigma(X_{0}^{\pi},\ldots,X_{n}^{\pi}).
\]
This definition seems, at first glance, circular... it seems as though
$\pi_{n}$ depends on $\mathcal{F}_{n}$ and vice versa. However,
if we look a bit closer at the definition of $X_{n}^{\pi}$, we note
that it only depends on the $\pi_{0},\ldots,\pi_{n-1}$. In light
of this, we can write $X_{n}^{\pi}\equiv X_{n}^{\pi_{0},\ldots,\pi_{n-1}}$.
Therefore,
\[
\mathcal{F}_{n}=\sigma(X_{0},X_{1}^{\pi_{0}},X_{2}^{\pi_{0},\pi_{1}},\ldots,X_{n}^{\pi_{0},\pi_{1},\ldots,\pi_{n-1}})
\]
and hence $\pi$ is well-defined.
\end{rem}
\section{Dynamic programming}
By the Markov property,
\begin{multline*}
J(i,\pi)=\mathbb{E}^{i}\left[c(X_{0}^{\pi})+\sum_{n\geq1}d^{n}c(X_{n}^{\pi})\right]=c(i)+d\mathbb{E}^{i}\left[\sum_{n\geq0}d^{n}c(X_{n+1}^{\pi})\right]\\
=c(i)+d\mathbb{E}^{i}\left[J(X_{1}^{\pi},(\pi_{n})_{n\geq1})\right]\geq c(i)+d\sum_{j}(P(\pi_{0}))_{ij}v(j)
\end{multline*}
where $\pi_{0}$ is some stationary policy. Taking infimums of both
sides of this equality,
\begin{equation}
v(i)\geq\inf_{\pi_{0}\in\Pi_{0}}\left\{ c(i)+d\sum_{j}(P(\pi_{0}))_{ij}v(j)\right\} .\label{eq:ldp}
\end{equation}
Now, fix $\epsilon>0$. For each $i$, let $\pi^{i}=(\pi_{n}^{i})_{n\geq0}$
be a randomized policy which satisifes
\[
v(i)\geq J(i,\pi^{i})-\epsilon.
\]
Let $\pi_{0}$ be an arbitrary stationary policy. Define a new randomized
policy $\pi^{\epsilon}=(\pi_{n}^{\epsilon})_{n\geq0}$ by
\[
\pi_{n}^{\epsilon}=\begin{cases}
\pi_{0} & \text{if }n=0\\
\sum_{i}\mathbf{1}_{\{X_{1}^{\pi_{0}}=i\}}\pi_{n-1}^{i} & \text{if }n>0.
\end{cases}
\]
Note that
\[
v(i)\leq J(i,\pi^{\epsilon})=c(i)+d\sum_{j}(P(\pi_{0}))_{ij}J(j,\pi^{j})\leq c(i)+d\sum_{j}(P(\pi_{0}))_{ij}v(j)+\epsilon.
\]
Now, take infimums of both sides to get
\begin{equation}
v(i)\leq\inf_{\pi_{0}\in\Pi_{0}}\left\{ c(i)+d\sum_{j}(P(\pi_{0}))_{ij}v(j)\right\} +\epsilon.\label{eq:udp}
\end{equation}
We can take $\epsilon\downarrow0$ and combine (\ref{eq:ldp}) and
(\ref{eq:udp}) to arrive at
\begin{equation}
v(i)=\inf_{\pi_{0}\in\Pi_{0}}\left\{ c(i)+d\sum_{j}(P(\pi_{0}))_{ij}v(j)\right\} .\label{eq:dp}
\end{equation}
The implications of this are amazing! We started out with an objective
function (\ref{eq:objective}) that was daunting: minimizing it would
require picking a stationary policy for each time $n$. However, we
were able to use the Markov property to reduce this to a ``local''
problem that only involves minimizing over all stationary policies
$\pi_{0}$. In fact, we can simplify (\ref{eq:dp}) even further.
First, we need some notation:
\[
\text{for }\{y_{\alpha}\}_{\alpha}\in\mathbb{R}^{n}\text{, }\inf_{\alpha}y_{\alpha}\text{ is the vector with entries }\inf_{\alpha}(y_{\alpha})_{i}.
\]
\begin{thm}[Dynamic programming]
Let $v=(v(1),\ldots,v(m))^{\intercal}$ and $c=(c(1),\ldots,c(m))^{\intercal}$
where $v(i)$ is the quantity defined by (\ref{eq:mdp}). Then,
\[
\boxed{\sup_{\pi_{0}\in\Pi_{0}}\left\{ \left(I-dP(\pi_{0})\right)v-c\right\} =0}
\]
\end{thm}
\begin{proof}
We can rewrite (\ref{eq:dp}) as
\begin{equation}
v=\inf_{\pi_{0}\in\Pi}\left\{ c+dP(\pi_{0})v\right\} .\label{eq:fixpoint}
\end{equation}
Moving some terms around, we obtain the desired result.
\end{proof}
In fact, the situation is much more general than we have let on. We
can allow for more general discount factors and costs:
\[
J(i,\pi)=\mathbb{E}\left[\sum_{n\geq0}d(\pi_{n}(X_{n}^{\pi}),X_{n}^{\pi})^{n}c(\pi_{n}(X_{n}^{\pi}),X_{n}^{\pi})\middle|X_{0}^{\pi}=i\right].
\]
However, in this case, it is no longer necessarily the case that $v(\cdot)$
is bounded. When it is, the corresponding dynamic programming equation
is
\begin{equation}
\sup_{\pi_{0}\in\Pi_{0}}\left\{ \left(I-D(\pi_{0})P(\pi_{0})\right)v-c(\pi_{0})\right\} =0\label{eq:dp_general}
\end{equation}
where $c(\pi_{0})=(c(\pi_{0}(1),1),\ldots,c(\pi_{0}(m),m))^{\intercal}$
and
\[
D(\pi_{0})=\operatorname{diag}(d(\pi_{0}(1),1),\ldots,d(\pi_{0}(m),m))=\begin{pmatrix}d(\pi_{0}(1),1)\\
& d(\pi_{0}(2),2)\\
& & \ddots\\
& & & d(\pi_{0}(m),m)
\end{pmatrix}.
\]
In light of this, the remainder of this lecture is focused on (\ref{eq:dp_general}).
In particular, we would like to know if an arbitrary vector $v$ satisfies
(\ref{eq:dp_general}), is it necessarily equal to the MDP (\ref{eq:mdp})?
Moreover, can we use (\ref{eq:dp_general}) to compute the MDP?
\end{document}