Markov Decision Processes

Deep Reinforcement Learning

2013 Atari (DQN)
[Deepmind]

Pong Enduro Beamrider Q*bert

 Deep Reinforcement Learning
2013 Atari (DQN)
[Deepmind]

2015 AlphaGo
[Deepmind]

AlphaGo Silver et al, Nature 2015

AlphaGoZero Silver et al, Nature 2017
AlphaZero Silver et al, 2017
Tian et al, 2016; Maddison et al, 2014; Clark et al, 2015
 Deep Reinforcement Learning
2013 Atari (DQN)
[Deepmind]

2015 AlphaGo
[Deepmind]

2016 3D locomotion (TRPO+GAE)

[Berkeley]

[Schulman, Moritz, Levine, Jordan, Abbeel, ICLR 2016]

 Deep Reinforcement Learning
2013 Atari (DQN)
[Deepmind]

2015 AlphaGo
[Deepmind]

2016 3D locomotion (TRPO+GAE)

[Berkeley]

2016 Real Robot Manipulation (GPS)

[Berkeley]

[Levine*, Finn*, Darrell, Abbeel, JMLR 2016]

 Deep Reinforcement Learning
2013 Atari (DQN)
[Deepmind]

2015 AlphaGo
[Deepmind]

2016 3D locomotion (TRPO+GAE)

[Berkeley]

2016 Real Robot Manipulation (GPS)

[Berkeley]

2019 Rubik’s Cube (PPO+DR) OpenAI

[OpenAI]
Non-Deterministic Search
 Example: Grid World
▪ A maze-like problem
▪ The agent lives in a grid
▪ Walls block the agent’s path

▪ Noisy movement: actions do not always go as

planned
▪ 80% of the time, the action North takes the agent
North (if there is no wall there)
▪ 10% of the time, North takes the agent West; 10% East
▪ If there is a wall in the direction the agent would have
been taken, the agent stays put

▪ The agent receives rewards

▪ Small “living” reward each step (can be negative)
▪ Big rewards come at the end (good or bad)

▪ Goal: maximize sum of rewards

 Grid World Actions
Deterministic Grid World Stochastic Grid World
 Markov Decision Processes
o An MDP is defined by:
o A set of states s  S
o A set of actions a  A
o A transition function T(s, a, s’)
o Probability that a from s leads to s’, i.e., P(s’| s, a)
o Also called the model or the dynamics
o A reward function R(s, a, s’)
o Sometimes just R(s) or R(s’)
o A start state
o Maybe a terminal state
 What is Markov about MDPs?
o “Markov” generally means that given the present state, the
future and the past are independent

o For Markov decision processes, “Markov” means action

outcomes depend only on the current state

Andrey Markov
(1856-1922)

o This is just like search, where the successor function could

only depend on the current state (not the history)
 Policies
o In deterministic single-agent search
problems, we wanted an optimal plan, or
sequence of actions, from start to a goal

o For MDPs, we want an optimal

policy *: S → A
o A policy  gives an action for each state
o An optimal policy is one that maximizes
expected utility if followed
Optimal policy when R(s, a, s’) = -0.03
o An explicit policy defines a reflex agent
for all non-terminals s
Optimal Policies
 Example: Racing
o A robot car wants to travel far, quickly
o Three states: Cool, Warm, Overheated
o Two actions: Slow, Fast
0.5 +1
o Going faster gets double reward
1.0
Fast
Slow -10
+1
0.5

Warm
Slow

Fast 0.5 +2

Cool 0.5
+1 Overheated
1.0
+2
Racing Search Tree
 MDP Search Trees
o Each MDP state projects an expectimax-like search tree

s s is a state

(s, a) is a q-
s, a
state
(s,a,s’) called a transition

s,a,s’ T(s,a,s’) = P(s’|s,a)

R(s,a,s’)
s’
Utilities of Sequences
 Utilities of Sequences
o What preferences should an agent have over reward sequences?

o More or less? [1, 2, 2] or [2, 3, 4]

o Now or later? [0, 0, 1] or [1, 0, 0]

 Discounting
o It’s reasonable to maximize the sum of rewards
o It’s also reasonable to prefer rewards now to rewards later
o One solution: values of rewards decay exponentially

Worth Now Worth Next Step Worth In Two Steps

 Discounting
o How to discount?
o Each time we descend a level, we
multiply in the discount once

o Why discount?
o Reward now is better than later
o Can also think of it as a 1-gamma
chance of ending the process at
every step
o Also helps our algorithms
converge

o Example: discount of 0.5

o U([1,2,3]) = 1*1 + 0.5*2 + 0.25*3
o U([1,2,3]) < U([3,2,1])
21
Compute Probability

22
23
24
25
26
27
28
29
30
31
Rewards and Utility

32
Rewards and Utility

33
Discounted Rewards and Utility

34
Stochastic World

35
 Stochastic World
o the utility in stochastic environment as

o where E is the expectation with respect to the probability

distribution over the possible paths following a certain policy.

36
 Stochastic World
o With stochastic, the utility of a state under a certain policy can
also be represented as the sum of its immediate reward and the
utility of its successor state, of course, following a probability
distribution.

37
38
Example: Racing
 Example: Racing
o A robot car wants to travel far, quickly
o Three states: Cool, Warm, Overheated
o Two actions: Slow, Fast
0.5 +1
o Going faster gets double reward
1.0
Fast
Slow -10
+1
0.5

Warm
Slow

Fast 0.5 +2

Cool 0.5
+1 Overheated
1.0
+2
Racing Search Tree
 MDP Search Trees
o Each MDP state projects an expectimax-like search tree

s s is a state

(s, a) is a q-
s, a
state
(s,a,s’) called a transition

s,a,s’ T(s,a,s’) = P(s’|s,a)

R(s,a,s’)
s’
Utilities of Sequences
 Utilities of Sequences
o What preferences should an agent have over reward sequences?

o More or less? [1, 2, 2] or [2, 3, 4]

o Now or later? [0, 0, 1] or [1, 0, 0]

 Discounting
o It’s reasonable to maximize the sum of rewards
o It’s also reasonable to prefer rewards now to rewards later
o One solution: values of rewards decay exponentially

Worth Now Worth Next Step Worth In Two Steps

 Discounting
o How to discount?
o Each time we descend a level, we
multiply in the discount once

o Why discount?
o Reward now is better than later
o Can also think of it as a 1-gamma
chance of ending the process at
every step
o Also helps our algorithms
converge

o Example: discount of 0.5

o U([1,2,3]) = 1*1 + 0.5*2 + 0.25*3
o U([1,2,3]) < U([3,2,1])
 Stationary Preferences
o Theorem: if we assume stationary preferences:

o Then: there are only two ways to define utilities

o Additive utility:
o Discounted utility:
 Quiz: Discounting
o Given:

o Actions: East, West, and Exit (only available in exit states a, e)

o Transitions: deterministic

o Quiz 1: For  = 1, what is the optimal policy? <- <- <-

o Quiz 2: For  = 0.1, what is the optimal policy? <- <- ->

o Quiz 3: For which  are West and East equally good when in state d?
10 3
 Infinite Utilities?!
▪ Problem: What if the game lasts forever? Do we get infinite
rewards?
▪ Solutions:
▪ Finite horizon: (similar to depth-limited search)
▪ Terminate episodes after a fixed T steps (e.g. life)
▪ Gives nonstationary policies ( depends on time left)

▪ Discounting: use 0 <  < 1

▪ Smaller  means smaller “horizon” – shorter term focus

▪ Absorbing state: guarantee that for every policy, a terminal state will
eventually be reached (like “overheated” for racing)
 Recap: Defining MDPs
o Markov decision processes: s
o Set of states S
o Start state s0 a
o Set of actions A s, a
o Transitions P(s’|s,a) (or T(s,a,s’))
o Rewards R(s,a,s’) (and discount ) s,a,s’
s’

o MDP quantities so far:

o Policy = Choice of action for each state
o Utility = sum of (discounted) rewards
Solving MDPs
 Recall: Racing MDP
o A robot car wants to travel far, quickly
o Three states: Cool, Warm, Overheated
o Two actions: Slow, Fast
0.5 +1
o Going faster gets double reward
1.0
Fast
Slow -10
+1
0.5

Warm
Slow

Fast 0.5 +2

Cool 0.5
+1 Overheated
1.0
+2
Racing Search Tree
Racing Search Tree
 Racing Search Tree
o We’re doing way too much
work with expectimax!

o Problem: States are repeated

o Idea: Only compute needed
quantities once

o Problem: Tree goes on

forever
o Idea: Do a depth-limited
computation, but with
increasing depths until change
is small
o Note: deep parts of the tree
eventually don’t matter if γ < 1
 Optimal Quantities

▪ The value (utility) of a state s:

V*(s) = expected utility starting in s s s is a
state
and acting optimally
a
(s, a) is a
▪ The value (utility) of a q-state (s,a): s, a q-state
Q*(s,a) = expected utility starting out
s,a,s’ (s,a,s’) is a
having taken action a from state s
transition
and (thereafter) acting optimally s’

▪ The optimal policy:

*(s) = optimal action from state s
Gridworld V* Values

Noise = 0.2
Discount = 0.9
Living reward = 0
Gridworld Q* Values

Noise = 0.2
Discount = 0.9
Living reward = 0
 Values of States (Bellman Eqns)
o Recursive definition of value:
s

a
s, a

s,a,s’
s’
Policy Evaluation

60
Policy Improvement

61
k=0

Noise = 0.2
Discount = 0.9
Living reward = 0
k=1

Noise = 0.2
Discount = 0.9
Living reward = 0
k=2

Noise = 0.2
Discount = 0.9
Living reward = 0
k=3

Noise = 0.2
Discount = 0.9
Living reward = 0
k=4

Noise = 0.2
Discount = 0.9
Living reward = 0
k=5

Noise = 0.2
Discount = 0.9
Living reward = 0
k=6

Noise = 0.2
Discount = 0.9
Living reward = 0
k=7

Noise = 0.2
Discount = 0.9
Living reward = 0
k=8

Noise = 0.2
Discount = 0.9
Living reward = 0
k=9

Noise = 0.2
Discount = 0.9
Living reward = 0
k=10

Noise = 0.2
Discount = 0.9
Living reward = 0
k=11

Noise = 0.2
Discount = 0.9
Living reward = 0
k=12

Noise = 0.2
Discount = 0.9
Living reward = 0
k=100

Noise = 0.2
Discount = 0.9
Living reward = 0
 Policy Iteration
o In policy iteration, we iteratively alternate policy evaluation and
policy improvement. In policy evaluation, we keep policy
constant and update utility based on that policy.

o In policy improvement, we keep utility constant and update

policy based on that utility.

76
 Policy Iteration
o The whole process of policy iteration is then

o We start with any arbitrary policy π₁, gets its utility v₁ by policy
evaluation, gets a new policy π₂ from v₁ by policy improvement,
gets utility v₂ of π₂ by policy evaluation, … until we converge on
our optimal policy π*.

77
 Time-Limited Values
o Key idea: time-limited values

o Define Vk(s) to be the optimal value of s if the game

ends in k more time steps
o Equivalently, it’s what a depth-k expectimax would give
from s
Computing Time-Limited Values
Value Iteration
 Value Iteration
o Start with V0(s) = 0: no time steps left means an expected reward sum of zero
o Given vector of Vk(s) values, do one ply of expectimax from each state:
Vk+1(s)
a
s, a

o Repeat until convergence, which yields V* s,a,s’

Vk(s’)

o Complexity of each iteration: O(S2A)

o Theorem: will converge to unique optimal values

o Basic idea: approximations get refined towards optimal values
o Policy may converge long before values do
82
83
84
85
86
87
88
 Convergence*
o How do we know the Vk vectors are going to
converge? (assuming 0 < γ < 1)

o Proof Sketch:
o For any state Vk and Vk+1 can be viewed as depth k+1
expectimax results in nearly identical search trees
o The difference is that on the bottom layer, Vk+1 has
actual rewards while Vk has zeros
o That last layer is at best all RMAX
o It is at worst RMIN
o But everything is discounted by γk that far out
o So Vk and Vk+1 are at most γk max|R| different
o So as k increases, the values converge