Reinforcement Learning : Overview

Reinforcement Learning: Overview

Cheng-Zhong XuWayne State University

Introduction In RL, the learner is a decision-making agent that takes actions in

an environment state and receives reward (or penalty) for its actions. The action may cause the change of environment state. After a set of trial-and-error runs, it should learn the best policy: the sequence of actions that maximize the total reward Supervised learning: learning from examples provided by a teacher RL: learning with a critic (reward or penalty); goal-directed learning

from interaction Examples:

Game-playing: Sequence of moves to win a game Robot in a maze: Sequence of actions to find a goal

2C. Xu, 2008

3

Example: K-armed Bandit Given $10 to play on a slot machine with 5 levers: Each play costs $1; each pull of a lever may produce payoff of 0, 1$, 5$, 10$ Find the optimal policy that pay off the most.

Tradeoff between exploitation and explorationExploitation: continue to pull the lever that returns positive Exploration: try to pull a new one

Deterministic model The payoff of each lever is fixed,

but unknown in advance Stochastic model The pay of each lever is uncertainty,

with known or unknown probabilityC. Xu, 2008

K-armed Bandit in General In deterministic case:

Q(a): value of action a

Reward of act a is ra

Q(a)= ra

Choose a* if

Q(a*)=maxa Q(a)

In stochastic model: Reward is non-deterministic: p(r|a)

Qt(a): estimate of the value of act a at time t

Delta Rule

is learning factor Qt+1(a) is expected value and should converge to the mean of p(r|a) as t increases

4

1 1.t t t tQ a Q a r a Q a

C. Xu, 2008

K-Armed Bandit as Simplified RL Single state (single slot machine) vs Multiple States

p(r|si , aj) : different reward probabilities

Q(Si aj ): value of action aj in state si to be learnt

Action causes state change, in addition to reward Rewards are not necessarily immediate value

Delayed rewards

5

Start S2

S3S4

S5 Goal

S7S8

C. Xu, 2008

6

Elements of RL

st : State of agent at time t

at: Action taken at time t

In st, action at is taken, clock ticks and reward rt+1 is received and state changes to st+1

Next state prob: P (st+1 | st , at ) Markov system

Reward prob: p (rt+1 | st , at ) Initial state(s), goal state(s) Episode (trial) of actions from initial state to goal

C. Xu, 2008

7

Policy and Cumulative Reward

Policy, State value of a policy, Finite-horizon:

Infinite horizon:

tt sa: AS tsV

T

iitTtttt rErrrEsV

121

rate discount the is 10

1

13

221

iit

itttt rErrrEsV

C. Xu, 2008

8

Bellman’s equation

1

11

* *1 1 1

*1 1

1

1

11 1

1

1 1

* *

[ ]

max ,

max | ,

, | ,

Value of in

max

tt

t

t

t

t t t t

it t i

i

it t i

i

t t

t t t t ta

t ta

s

t t t t t ta

s

V s E r

E r r

V s E r P s s a V s

Q s a E r

E r E V s

V s Q s a a

P s

s

s a

*1 1,t tQ s a

State Value Function Example GridWorld: a simple MDP

Grid cell ~ environment states Four possible actions at each cell: n/s/e/w, one cell in

respective dir; Agent would remain in location, if its move would take it off

the grid, but with reward of -1; Other move receives reward of 0, except

Those moves out of states A and B; rewarding 10 for each move out of A (to A’) and 5 for move out of B (to B’)

Policy: the agent selects four actions with equal prob and assume =0.9

9

10

Environment, P (st+1 | st , at ), p (rt+1 | st , at ), is known

There is no need for exploration Can be solved using dynamic programming Solve for

Optimal policy

Model-Based Learning

111

1

|max t*

stttt

at

* sVa,ssPrEsVt

t

111

1

||max arg t*

stttttt

at sVa,ssPa,srEs*

tt

C. Xu, 2008

11

Value Iteration vs Policy Iteration

Policy iteration needs fewer iterations than value iteration

C. Xu, 2008

12

Model-Free Learning

Environment, P (st+1 | st , at ), p (rt+1 | st , at ), is not known model-free learning, based on both exploitation and exploration

Temporal difference learning: use the (discounted) reward received in the next time step to update the value of current state (action): 1-step TD Temporal difference: between the value of the

current action and the value discounted from the next state

C. Xu, 2008

13

Deterministic Rewards and Actions

is reduced to

Therefore, we have a backup update rule as

Initially, and its value increases as learning proceeds episode by episode.

11111

1

max|

tt*

as

tttttt* a,sQa,ssPrEa,sQ

tt

1111

max

tta

ttt a,sQra,sQt

1111

max

tta

ttt a,sQ̂ra,sQ̂t

ˆ , 0,t tQ s a

Start S2

S3S4

S5 Goal

S7S8

In maze, all rewards of intermediate states are zero in the first episode. We a goal is reached, we get reward r and the Q value of last state, say S5, is Updated as r. In the next episode, when S5 is reached,the Q value of its preceding state S4 is updated as 2r.

C. Xu, 2008

14

Nondeterministic Rewards and Actions

Uncertainty in reward and state change is due To presence of opponents or randomness in the environment.

Q-learning (Watkins & Dayan’92): we keep a running average for each pair of state-action

tttt

attttt a,sQ̂a,sQ̂ra,sQ̂a,sQ̂

t111

1

max

value of a sample of instances for each (st,at)

11111

1

max|

tt*

as

tttttt* a,sQa,ssPrEa,sQ

tt

C. Xu, 2008

15

Exploration Strategies

Greedy: choose action that maximizes the immediate reward

ε-greedy: with prob ε, choose one action at random uniformly, and choose the best action with prob 1-ε

Softmax selection:

To m gradually move from exploration to exploitation, temperature variable T could help the annealing process

,

,

1

e|

e

Q s a

Q s b

b

P a s

A

( , ) /

( , ) /

1

e|

e

Q s a T

Q s b T

b

P a s

A

C. Xu, 2008

Summary RL is a process of learning by interaction, in contrast to

supervised learning from examples. Elements of RL for an agent and its environment

state value function, state-action function (Q-value), reward, state change probability, policy

Tradeoff between exploitation and exploration Markov Decision Process Model-based learning

Value function in Bellman equation Dynamic programming

Model-free learning Temporal difference (TD) and Q-learning (timing average) to update

Q value Action selection for exploration

-greedy, softmax-based selection

16C. Xu, 2008