CS 188: Artificial Intelligence

Markov Decision Processes II

Instructor: Pieter Abbeel

University of California, Berkeley
[These slides adapted from Dan Klein and Pieter Abbeel]
 Recap: Defining MDPs
o Markov decision processes:
s
o Set of states S
o Start state s0 a
o Set of actions A
o Transitions P(s’|s,a) (or T(s,a,s’)) s, a
o Rewards R(s,a,s’) (and discount g)
s,a,s’
s’
o MDP quantities so far:
o Policy = Choice of action for each state
o Utility = sum of (discounted) rewards
o Values = expected future utility from a state (max node)
o Q-Values = expected future utility from a q-state (chance node)
 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

Solving MDPs
 Optimal Quantities

§ The value (utility) of a state s:

V*(s) = expected utility starting in s s s is a
and acting optimally state
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
having taken action a from state s s,a,s’ (s,a,s’) is a
and (thereafter) acting optimally s’ transition

§ The optimal policy:

p*(s) = optimal action from state s
 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
 Value Iteration
o Bellman equations characterize the optimal values: V(s)

a
s, a

s,a,s’
o Value iteration computes them: V(s’)

o Value iteration is just a fixed point solution method

o … though the Vk vectors are also interpretable as time-limited values
 Value Iteration (again J ) s
a
o Init:
s, a
∀𝑠: 𝑉 𝑠 = 0
s,a,s’
o Iterate: s’

∀𝑠: 𝑉!"# 𝑠 = max * 𝑇 𝑠, 𝑎, 𝑠 & [𝑅 𝑠, 𝑎, 𝑠 & + 𝛾𝑉 𝑠 & ]

$
%&

𝑉 = 𝑉!"#
Note: can even directly assign to V(s), which will not compute the sequence of Vk but will still converge to V*
The Bellman Equations

How to be optimal:
Step 1: Take correct first action

Step 2: Keep being optimal

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 Extraction
 Computing Actions from Values
o Let’s imagine we have the optimal values V*(s)

o How should we act?

o It’s not obvious!

o We need to do a mini-expectimax (one step)

o This is called policy extraction, since it gets the policy implied by the
values
 Computing Actions from Q-Values
o Let’s imagine we have the optimal
q-values:

o How should we act?

o Completely trivial to decide!

o Important lesson: actions are easier to select from q-values than

values!
 Let’s think.
o Take a minute, think about value iteration.
o Write down the biggest question you have about it.
Policy Methods
 Problems with Value Iteration
o Value iteration repeats the Bellman updates: s
a
s, a

o Problem 1: It’s slow – O(S2A) per iteration s,a,s’

s’

o Problem 2: The “max” at each state rarely changes

o Problem 3: The policy often converges long before the values

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 Alternative approach for optimal values:
o Step 1: Policy Evaluation: calculate utilities for some fixed policy (not optimal
utilities!) until convergence
o Step 2: Policy Improvement: update policy using one-step look-ahead with
resulting converged (but not optimal!) utilities as future values
o Repeat steps until policy converges

o This is Policy Iteration

o It’s still optimal!
o Can converge (much) faster under some conditions
Policy Evaluation
 Fixed Policies
Do the optimal action Do what p says to do
s s
a p(s)
s, a s, p(s)

s,a,s’ s, p(s),s’
s’ s’

o Expectimax trees max over all actions to compute the optimal values
o If we fixed some policy p(s), then the tree would be simpler – only one action
per state
o … though the tree’s value would depend on which policy we fixed
 Utilities for a Fixed Policy
o Another basic operation: compute the utility of a state s s
under a fixed (generally non-optimal) policy
p(s)
o Define the utility of a state s, under a fixed policy p: s, p(s)
Vp(s) = expected total discounted rewards starting in s and
following p
s, p(s),s’
s’
o Recursive relation (one-step look-ahead / Bellman
equation):
 Policy Evaluation
o How do we calculate the V’s for a fixed policy p? s

o Idea 1: Turn recursive Bellman equations into updates p(s)

(like value iteration) s, p(s)

s, p(s),s’
s’

o Efficiency: O(S2) per iteration

o Idea 2: Without the maxes, the Bellman equations are just a linear system
o Solve with Matlab (or your favorite linear system solver)
 Example: Policy Evaluation
Always Go Right Always Go Forward
 Example: Policy Evaluation
Always Go Right Always Go Forward
Policy Iteration
 Policy Iteration

o Evaluation: For fixed current policy p, find values with policy evaluation:
o Iterate until values converge:

o Improvement: For fixed values, get a better policy using policy extraction
o One-step look-ahead:
 Comparison
o Both value iteration and policy iteration compute the same thing (all optimal values)

o In value iteration:
o Every iteration updates both the values and (implicitly) the policy
o We don’t track the policy, but taking the max over actions implicitly recomputes it

o In policy iteration:
o We do several passes that update utilities with fixed policy (each pass is fast because we
consider only one action, not all of them)
o After the policy is evaluated, a new policy is chosen (slow like a value iteration pass)
o The new policy will be better (or we’re done)

o Both are dynamic programs for solving MDPs

 Summary: MDP Algorithms
o So you want to….
o Compute optimal values: use value iteration or policy iteration
o Compute values for a particular policy: use policy evaluation
o Turn your values into a policy: use policy extraction (one-step lookahead)

o These all look the same!

o They basically are – they are all variations of Bellman updates
o They all use one-step lookahead expectimax fragments
o They differ only in whether we plug in a fixed policy or max over actions
The Bellman Equations

How to be optimal:
Step 1: Take correct first action

Step 2: Keep being optimal

Double Bandits
 Double-Bandit MDP
o Actions: Blue, Red No discount
o States: Win, Lose 100 time steps
0.25 $0
Both states have
the same value
0.75
$2
W 0.25 L
$0
$1 $1
0.75 $2
1.0 1.0
 Offline Planning
o Solving MDPs is offline planning No discount
o You determine all quantities through computation 100 time steps
o You need to know the details of the MDP Both states have
the same value
o You do not actually play the game!

0.25 $0
Value
0.75
W $2 0.25 L
Play Red 150 $0
$1 $1
0.75 $2
Play Blue 100 1.0 1.0
Let’s Play!

$2 $2 $0 $2 $2
$2 $2 $0 $0 $0
 Online Planning
o Rules changed! Red’s win chance is different.

?? $0

??
$2
W ?? L
$0
$1 $1
?? $2
1.0 1.0
Let’s Play!

$0 $0 $0 $2 $0
$2 $0 $0 $0 $0
 What Just Happened?
o That wasn’t planning, it was learning!
o Specifically, reinforcement learning
o There was an MDP, but you couldn’t solve it with just computation
o You needed to actually act to figure it out

o Important ideas in reinforcement learning that came up

o Exploration: you have to try unknown actions to get information
o Exploitation: eventually, you have to use what you know
o Regret: even if you learn intelligently, you make mistakes
o Sampling: because of chance, you have to try things repeatedly
o Difficulty: learning can be much harder than solving a known MDP
Next Time: Reinforcement Learning!