Artificial Intelligence Non-Deterministic Search

Markov Decision Processes

Example: Grid World Grid World Actions

Deterministic Grid World Stochastic 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 each time step

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

§ Goal: maximize sum of rewards

Markov Decision Processes What is Markov about MDPs?

§ An MDP is defined by: § “Markov” generally means that given the present state, the
§ A set of states s  S future and the past are independent
§ A set of actions a  A
§ A transition function T(s, a, s’)
§ Probability that a from s leads to s’, i.e., P(s’| s, a) § For Markov decision processes, “Markov” means action
§ Also called the model or the dynamics outcomes depend only on the current state
§ A reward function R(s, a, s’)
§ Sometimes just R(s) or R(s’)
§ A start state
§ Maybe a terminal state
Andrey Markov
(1856-1922)
§ MDPs are non-deterministic search problems
§ One way to solve them is with expectimax search § This is just like search, where the successor function could only
§ We’ll have a new tool soon
depend on the current state (not the history)

[Demo – gridworld manual intro (L8D1)]

1
 Policies Video of Demo Gridworld Manual Intro
§ In deterministic single-agent search problems,
we wanted an optimal plan, or sequence of
actions, from start to a goal

§ For MDPs, we want an optimal policy *: S → A

§ A policy  gives an action for each state
§ An optimal policy is one that maximizes
expected utility if followed
§ An explicit policy defines a reflex agent
Optimal policy when R(s, a, s’) = -0.03
for all non-terminals s
§ Expectimax didn’t compute entire policies
§ It computed the action for a single state only

Example: Racing Example: Racing

§ A robot car wants to travel far, quickly
§ Three states: Cool, Warm, Overheated
§ Two actions: Slow, Fast
0.5 +1
§ 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

§ Each MDP state projects an expectimax-like search tree

s s is a state

(s, a) is a
s, a
q-state
(s,a,s’) called a transition
s,a,s’ T(s,a,s’) = P(s’|s,a)
R(s,a,s’)
s’

2
 Utilities of Sequences Utilities of Sequences
§ What preferences should an agent have over reward sequences?

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

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

Discounting Discounting
§ It’s reasonable to maximize the sum of rewards § How to discount?
§ It’s also reasonable to prefer rewards now to rewards later § Each time we descend a level, we
multiply in the discount once
§ One solution: values of rewards decay exponentially
§ Why discount?
§ Sooner rewards probably do have
higher utility than later rewards
§ Also helps our algorithms converge

§ Example: discount of 0.5

§ U([1,2,3]) = 1*1 + 0.5*2 + 0.25*3
Worth Now Worth Next Step Worth In Two Steps
§ U([1,2,3]) < U([3,2,1])

Stationary Preferences Quiz: Discounting

§ Theorem: if we assume stationary preferences: § Given:

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

§ Transitions: deterministic

§ Quiz 1: For  = 1, what is the optimal policy?

§ Then: there are only two ways to define utilities
§ Quiz 2: For  = 0.1, what is the optimal policy?
§ Additive utility:
§ Discounted utility: § Quiz 3: For which  are West and East equally good when in state d?

3
 Infinite Utilities?! Recap: Defining MDPs
§ Problem: What if the game lasts forever? Do we get infinite rewards? § Markov decision processes: s
§ Solutions: § Set of states S
§ Start state s0 a
§ Finite horizon: (similar to depth-limited search)
§ Terminate episodes after a fixed T steps (e.g. life) § Set of actions A s, a
§ Gives nonstationary policies ( depends on time left) § Transitions P(s’|s,a) (or T(s,a,s’))
§ Rewards R(s,a,s’) (and discount ) s,a,s’
§ Discounting: use 0 <  < 1 s’

§ MDP quantities so far:

§ Smaller  means smaller “horizon” – shorter term focus § Policy = Choice of action for each state
§ Absorbing state: guarantee that for every policy, a terminal state will eventually § Utility = sum of (discounted) rewards
be reached (like “overheated” for racing)

Solving MDPs Optimal Quantities

§ The value (utility) of a state s:

V*(s) = expected utility starting in s and s s is a
state
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 and
transition
(thereafter) acting optimally s’

§ The optimal policy:

*(s) = optimal action from state s
[Demo – gridworld values (L8D4)]

Snapshot of Demo – Gridworld V Values Snapshot of Demo – Gridworld Q Values

Noise = 0 Noise = 0
Discount = 1 Discount = 1
Living reward = 0 Living reward = 0

4
 Snapshot of Demo – Gridworld V Values Snapshot of Demo – Gridworld Q Values

Noise = 0.2 Noise = 0.2

Discount = 1 Discount = 1
Living reward = 0 Living reward = 0

Snapshot of Demo – Gridworld V Values Snapshot of Demo – Gridworld Q Values

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

Values of States Racing Search Tree

§ Fundamental operation: compute the (expectimax) value of a state
§ Expected utility under optimal action
s
§ Average sum of (discounted) rewards
§ This is just what expectimax computed! a
s, a
§ Recursive definition of value:
s,a,s’
s’

5
 Racing Search Tree Racing Search Tree
§ We’re doing too much work
with expectimax!

§ Problem: States are repeated

§ Idea: Only compute needed
quantities once

§ Problem: Tree goes on forever

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

Time-Limited Values k=0

§ Key idea: time-limited values

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

in k more time steps
§ Equivalently, it’s what a depth-k expectimax would give from s

Noise = 0.2
Discount = 0.9
[Demo – time-limited values (L8D6)] Living reward = 0

k=1 k=2

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

6
k=3 k=4

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

k=5 k=6

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

k=7 k=8

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

7
 k=9 k=10

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

k=11 k=12

Noise = 0.2 Noise = 0.2

Discount = 0.9 Discount = 0.9
Living reward = 0 Living reward = 0

k=100 Computing Time-Limited Values

Noise = 0.2
Discount = 0.9
Living reward = 0

8
 Value Iteration Value Iteration
§ Start with V0(s) = 0: no time steps left means an expected reward sum of zero
§ Given vector of Vk(s) values, do one ply of expectimax from each state:
Vk+1(s)
a
s, a

§ Repeat until convergence s,a,s’

Vk(s’)

§ Complexity of each iteration: O(S2A)

§ Theorem: will converge to unique optimal values

§ Basic idea: approximations get refined towards optimal values
§ Policy may converge long before values do

Example: Value Iteration Convergence*

§ How do we know the Vk vectors are going to converge?

§ Case 1: If the tree has maximum depth M, then VM holds

3.5 2.5 0 the actual untruncated values

§ Case 2: If the discount is less than 1

§ Sketch: For any state Vk and Vk+1 can be viewed as depth
k+1 expectimax results in nearly identical search trees
2 1 0 § The difference is that on the bottom layer, Vk+1 has actual
rewards while Vk has zeros
Assume no discount! § That last layer is at best all RMAX
§ It is at worst RMIN
0 0 0 § But everything is discounted by γk that far out
§ So Vk and Vk+1 are at most γk max|R| different
§ So as k increases, the values converge

Next Time: Policy-Based Methods