R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 1 From Sutton & Barto Reinforcement Learning An Introduction

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 2 DP Value Iteration Recall the full policy-evaluation backup: Here is the full value-iteration backup:

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 3 Asynchronous DP pAll the DP methods described so far require exhaustive sweeps of the entire state set. pAsynchronous DP does not use sweeps. Instead it works like this: Repeat until convergence criterion is met: –Pick a state at random and apply the appropriate backup pStill need lots of computation, but does not get locked into hopelessly long sweeps pCan you select states to backup intelligently? YES: an agent’s experience can act as a guide.

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 4 Generalized Policy Iteration Generalized Policy Iteration (GPI): any interaction of policy evaluation and policy improvement, independent of their granularity. A geometric metaphor for convergence of GPI:

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 5 Efficiency of DP pTo find an optimal policy is polynomial in the number of states… pBUT, the number of states is often astronomical, e.g., often growing exponentially with the number of state variables (what Bellman called “the curse of dimensionality”). pIn practice, classical DP can be applied to problems with a few millions of states. pAsynchronous DP can be applied to larger problems, and appropriate for parallel computation. pIt is surprisingly easy to come up with MDPs for which DP methods are not practical.

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 6 DP - Summary pPolicy evaluation: backups without a max pPolicy improvement: form a greedy policy, if only locally pPolicy iteration: alternate the above two processes pValue iteration: backups with a max pFull backups (to be contrasted later with sample backups) pGeneralized Policy Iteration (GPI) pAsynchronous DP: a way to avoid exhaustive sweeps pBootstrapping: updating estimates based on other estimates

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 7 Chapter 5: Monte Carlo Methods pMonte Carlo methods learn from complete sample returns Only defined for episodic tasks pMonte Carlo methods learn directly from experience On-line: No model necessary and still attains optimality Simulated: No need for a full model

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 8 Monte Carlo Policy Evaluation  Goal: learn V  (s)  Given: some number of episodes under  which contain s pIdea: Average returns observed after visits to s pEvery-Visit MC: average returns for every time s is visited in an episode pFirst-visit MC: average returns only for first time s is visited in an episode pBoth converge asymptotically 12345

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 9 First-visit Monte Carlo policy evaluation

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 10 Blackjack example pObject: Have your card sum be greater than the dealers without exceeding 21. pStates (200 of them): current sum (12-21) dealer’s showing card (ace-10) do I have a useable ace? pReward: +1 for winning, 0 for a draw, -1 for losing pActions: stick (stop receiving cards), hit (receive another card) pPolicy: Stick if my sum is 20 or 21, else hit

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 11 Blackjack value functions

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 12 Backup diagram for Monte Carlo pEntire episode included pOnly one choice at each state (unlike DP) pMC does not bootstrap pTime required to estimate one state does not depend on the total number of states

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 13 Monte Carlo Estimation of Action Values (Q) pMonte Carlo is most useful when a model is not available We want to learn Q *  Q  (s,a) - average return starting from state s and action a following  pAlso converges asymptotically if every state-action pair is visited pExploring starts: Every state-action pair has a non-zero probability of being the starting pair

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 14 Monte Carlo Control pMC policy iteration: Policy evaluation using MC methods followed by policy improvement pPolicy improvement step: greedify with respect to value (or action-value) function

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 15 Convergence of MC Control pGreedified policy meets the conditions for policy improvement: pAnd thus must be ≥  k by the policy improvement theorem pThis assumes exploring starts and infinite number of episodes for MC policy evaluation pTo solve the latter: update only to a given level of performance alternate between evaluation and improvement per episode

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 16 Monte Carlo Exploring Starts Fixed point is optimal policy  * Now proven (almost)

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 17 Blackjack example continued pExploring starts pInitial policy as described before

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 18 On-policy Monte Carlo Control greedy non-max pOn-policy: learn about policy currently executing pHow do we get rid of exploring starts? Need soft policies:  (s,a) > 0 for all s and a e.g.  -soft policy:  Similar to GPI: move policy towards greedy policy (i.e.  - soft)  Converges to best  -soft policy

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 19 On-policy MC Control

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 20 Off-policy Monte Carlo control pBehavior policy generates behavior in environment pEstimation policy is policy being learned about pAverage returns from behavior policy by probability their probabilities in the estimation policy

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 21 Learning about  while following

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 22 Off-policy MC control

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 23 Incremental Implementation pMC can be implemented incrementally saves memory pCompute the weighted average of each return incremental equivalentnon-incremental

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 24 MC - Summary pMC has several advantages over DP: Can learn directly from interaction with environment No need for full models No need to learn about ALL states Less harm by Markovian violations (later in book) pMC methods provide an alternate policy evaluation process pOne issue to watch for: maintaining sufficient exploration exploring starts, soft policies pIntroduced distinction between on-policy and off-policy methods pNo bootstrapping (as opposed to DP)

R. S. Sutton and A. G. Barto: Reinforcement Learning: An Introduction 25 Monte Carlo is important in practice pAbsolutely pWhen there are just a few possibilities to value, out of a large state space, Monte Carlo is a big win pBackgammon, Go, …