Reinforcement Learning. George Konidaris

Size: px

Start display at page:

Download "Reinforcement Learning. George Konidaris"

Cordelia Oliver
5 years ago
Views:

1 Reinforcement Learning George Konidaris Fall 2017

2 Machine Learning Subfield of AI concerned with learning from data. Broadly, using: Experience To Improve Performance On Some Task (Tom Mitchell, 1997)

3 vs ML vs Statistics vs Data Mining

4 Why? Developing effective learning methods has proved difficult. Why bother? Autonomous discovery We don t know something, want to find out. Hard to program Easier to specify task, collect data. Adaptive behavior Our agents should adapt to new data, unforeseen circumstances.

5 Types of Machine Learning Depends on feedback available: Labeled data: Supervised learning No feedback, just data: Unsupervised learning. Sequential data, weak labels: Reinforcement learning

6 Supervised Learning Input: X = {x1,, xn} Y = {y1,, yn} inputs labels training data Learn to predict new labels. Given x: y?

7 Unsupervised Learning Input: X = {x1,, xn} inputs Try to understand the structure of the data. E.g., how many types of cars? How can they vary?

8 Reinforcement Learning Learning counterpart of planning. π : S A max R = t=0 t r t

9 MDPs Agent interacts with an environment At each time t: Receives sensor signal Executes action Transition: a t new sensor signal reward r t s t s t+1 Goal: find policy of discounted future rewards): [ π that maximizes expected return (sum max π E R = ] γ t r t t=0

10 Markov Decision Processes S A γ : set of states : set of actions : discount factor <S,A,γ,R,T > R : reward function R(s, a, s ) is the reward received taking action s and transitioning to state. a from state s T : transition function T (s s, a) is the probability of transitioning to state s taking action a in state s. after

11 RL vs Planning In planning: Transition function (T) known. Reward function (R) known. Computation offline. In reinforcement learning: One or both of T, R unknown. Action in the world only source of data. Transitions are executed not simulated.

12 Reinforcement Learning

13 RL This formulation is general enough to encompass a wide variety of learned control problems.

14 MDPs As before, our target is a policy: π : S A A policy maps states to actions. The optimal policy maximizes: [ ] max π s, E R(s) = t=0 γ t r t s 0 = s This means that we wish to find a policy that maximizes the return from every state.

15 Planning via Policy Iteration In planning, we used policy iteration to find an optimal policy. 1. Start with a policy V π 2. Estimate 3. Improve π a. (s) = max E [r + a V (s 0 )], 8s Repeat More precisely, we use a value function: V (s) =E " 1 X i=0 i r i # can t do this anymore then we would update π X (s) = argmax a by computing: s 0 T (s, a, s 0 )[r(s, a, s 0 )+ V [s 0 ]]

16 Value Functions For learning, we use a state-action value function as follows: " # X 1 Q (s, a) =E i=0 i r i s 0 = s, a 0 = a This is the value of executing ain state s, then following π. Note that V (s) =Q (s, (s)). A x

17 Policy Iteration This leads to a general policy improvement framework: 1. Start with a policy 2. Learn 3. Improve a. Q π π(s) = max a π Q(s, a), s Repeat Steps 2 and 3 can be interleaved as rapidly as you like. Usually, perform 3a every time step.

18 Value Function Learning Learning proceeds by gathering samples of Q(s, a). Methods differ by: How you get the samples. How you use them to update Q.

19 Monte Carlo Simplest thing you can do: sample R(s). r r r r r r r r Do this repeatedly, average values: Q(s, a) = R 1(s)+R 2 (s) R n (s) n

20 Temporal Difference Learning Where can we get more (immediate) samples? Idea: use the Bellman equation. V (s) =E s 0 [r(s, (s),s 0 )+ V (s 0 )] reward value of this state value of next state

21 TD Learning Ideally and in expectation: r i + Q(s i+1,a i+1 ) Q(s i,a i )=0 Q is correct if this holds in expectation for all states. When it does not: temporal difference error. s t s t+1 a t r t Q(s t,a t ) r t + γq(s t+1,a t+1 )

22 Sarsa Sarsa: very simple algorithm 1. Initialize Q[s][a] = 0 2. For n episodes observe state s select a = argmaxa Q[s][a] observe transition compute TD error (s, a, r, s,a ) δ = r + γq(s,a ) Q(s, a) update Q: Q(s, a) =Q(s, a)+αδ if not end of episode, repeat zero by def. if s is absorbing

23 Sarsa

24 Sarsa

25 Exploration vs. Exploitation Always maxa Q(s, a)? Exploit current knowledge. What if your current knowledge is wrong? How are you going to find out? Explore to gain new knowledge. Exploration vs. Exploitation - when to try new things? Consistent theme of RL.

26 Exploration vs. Exploitation How to balance? Simplest, most popular approach: Instead of always being greedy: maxa Q(s, a) Explore with probability : maxa Q(s, a) with probability (1 ). random action with probability. ( 0.1) - greedy exploration Very simple Ensures asymptotic coverage of state space

27 TD vs. MC TD and MC two extremes of obtaining samples of Q: r + γv r + γv r + γv... t=1 t=2 t=3 t=4 t=l γ i r i... t=1 t=2 t=3 t=4 t=l i

28 Generalizing TD We can generalize this to the idea of an n-step rollout: Each tells us something about the value function. We can combine all n-step rollouts. This is known as a complex backup.

29 TD(λ) Weighted sum: R (1) = r 0 + V (s 1 ) R (2) = r. 0 + r V (s 2 ).. R (n) = n 1 X i=0 i r i + n V (s n ) 1 n weights Estimator:

30 Sarsa(λ) This is called the λ-return. At λ=0 we get Sarsa, at λ=1 we get MC. Intermediate values of λ usually best. TD(λ) family of algorithms

31 Sarsa(λ): Implementation Each state has eligibility trace e(s, a). At time t: e(st, at) = 1 e(s, a) = γλe(s, a), for all other (s, a) pairs. At end of episode: e(s, a) = 0, for all (s, a) pairs. When updating: Compute δ as before Q(s, a) = Q(s, a) + αδe(s, a), for each (s, a) pair.

32 Sarsa(λ): Implementation 1. Initialize Q[s][a] = 0, for all (s, a) 2. Initialize e[s][a] = 0, for all (s, a) 2. For n episodes observe state st select at = argmaxa Q(st, a) observe transition (st, at, rt, st+1, at+1) compute TD error = r t + Q(s t+1,a t+1 ) Q(s t,a t ) e(st, at) = 1; other e(s, a) = γλe(s, a) update Q: Q(s, a) =Q(s, a)+ e(s, a) if not end of episode, repeat if end of episode, e[s][a] = 0 for all (s,a) don t forget! zero by def. if s is absorbing

33 Sarsa(λ)

34 Sarsa(λ)

35 Next Week: More Realism

Temporal Difference. Learning KENNETH TRAN. Principal Research Engineer, MSR AI

Temporal Difference. Learning KENNETH TRAN. Principal Research Engineer, MSR AI Temporal Difference Learning KENNETH TRAN Principal Research Engineer, MSR AI Temporal Difference Learning Policy Evaluation Intro to model-free learning Monte Carlo Learning Temporal Difference Learning