CS 188: Artificial Intelligence Spring 2007 Lecture 21:Reinforcement Learning: II MDP 4/12/2007 Srini Narayanan – ICSI and UC Berkeley Announcements Othello tournament signup Please send email to cs188@imail.berkeley.edu HW on classification out Due 4/23 Can work in pairs Reinforcement Learning Basic idea: Receive feedback in the form of rewards Agent’s utility is defined by the reward function Must learn to act so as to maximize expected utility Change the rewards, change the behavior Examples: Learning optimal paths Playing a game, reward at the end for winning / losing Vacuuming a house, reward for each piece of dirt picked up Automated taxi, reward for each passenger delivered Recap: MDPs Markov decision processes: States S Actions A Transitions P(s’|s,a) (or T(s,a,s’)) Rewards R(s,a,s’) Start state s0 Examples: Gridworld, High-Low, N-Armed Bandit Any process where the result of your action is stochastic Goal: find the “best” policy Policies are maps from states to actions What do we mean by “best”? This is like search – it’s planning using a model, not actually interacting with the environment MDP Solutions In deterministic single-agent search, want an optimal sequence of actions from start to a goal In an MDP, like expectimax, want an optimal policy (s) A policy gives an action for each state Optimal policy maximizes expected utility (i.e. expected rewards) if followed Defines a reflex agent Optimal policy when R(s, a, s’) = -0.04 for all non-terminals s Example Optimal Policies R(s) = -0.01 R(s) = -0.03 R(s) = -0.4 R(s) = -2.0 Stationarity In order to formalize optimality of a policy, need to understand utilities of reward sequences Typically consider stationary preferences: Theorem: only two ways to define stationary utilities Additive utility: Discounted utility: Infinite Utilities?! Problem: infinite state sequences with infinite rewards Solutions: Finite horizon: Terminate after a fixed T steps Gives nonstationary policy ( depends on time left) Absorbing state(s): guarantee that for every policy, agent will eventually “die” (like “done” for High-Low) Discounting: for 0 < < 1 Smaller means smaller horizon How (Not) to Solve an MDP The inefficient way: Enumerate policies For each one, calculate the expected utility (discounted rewards) from the start state E.g. by simulating a bunch of runs Choose the best policy We’ll return to a (better) idea like this later Utility of a State Define the utility of a state under a policy: V(s) = expected total (discounted) rewards starting in s and following Recursive definition (one-step look-ahead): Policy Evaluation Idea one: turn recursive equations into updates Idea two: it’s just a linear system, solve with Matlab (or Mosek, or Cplex) Example: High-Low Policy: always say “high” Iterative updates: Optimal Utilities Goal: calculate the optimal utility of each state V*(s) = expected (discounted) rewards with optimal actions Why: Given optimal utilities, MEU tells us the optimal policy Bellman’s Equation for Selecting actions Definition of utility leads to a simple relationship amongst optimal utility values: Optimal rewards = maximize over first action and then follow optimal policy Formally: Bellman’s Equation That’s my equation! Example: GridWorld Value Iteration Idea: Start with bad guesses at all utility values (e.g. V0(s) = 0) Update all values simultaneously using the Bellman equation (called a value update or Bellman update): Repeat until convergence Theorem: will converge to unique optimal values Basic idea: bad guesses get refined towards optimal values Policy may converge long before values do Example: Bellman Updates Example: Value Iteration Information propagates outward from terminal states and eventually all states have correct value estimates [DEMO] Convergence* Define the max-norm: Theorem: For any two approximations U and V (any two utility vectors) I.e. any distinct approximations must get closer to each other (after the Bellman update), so, in particular, any approximation must get closer to the true U (Bellman update is U) and value iteration converges to a unique, stable, optimal solution Theorem: I.e. once the change in our approximation is small, it must also be close to correct Policy Iteration Alternate approach: Policy evaluation: calculate utilities for a fixed policy until convergence (remember the beginning of lecture) Policy improvement: update policy based on resulting converged utilities Repeat until policy converges This is policy iteration Can converge faster under some conditions Policy Iteration If we have a fixed policy , use simplified Bellman equation to calculate utilities: For fixed utilities, easy to find the best action according to one-step look-ahead Comparison In value iteration: Every pass (or “backup”) updates both utilities (explicitly, based on current utilities) and policy (possibly implicitly, based on current policy) In policy iteration: Several passes to update utilities with frozen policy Occasional passes to update policies Hybrid approaches (asynchronous policy iteration): Any sequences of partial updates to either policy entries or utilities will converge if every state is visited infinitely often Reinforcement Learning Reinforcement learning: Still have an MDP: A set of states s S A model T(s,a,s’) A reward function R(s) Still looking for a policy (s) New twist: don’t know T or R I.e. don’t know which states are good or what the actions do Must actually try actions and states out to learn Example: Animal Learning RL studied experimentally for more than 60 years in psychology Rewards: food, pain, hunger, drugs, etc. Mechanisms and sophistication debated Example: foraging Bees learn near-optimal foraging plan in field of artificial flowers with controlled nectar supplies Bees have a direct neural connection from nectar intake measurement to motor planning area Example: Backgammon Reward only for win / loss in terminal states, zero otherwise TD-Gammon learns a function approximation to U(s) using a neural network Combined with depth 3 search, one of the top 3 players in the world Passive Learning Simplified task You don’t know the transitions T(s,a,s’) You don’t know the rewards R(s,a,s’) You are given a policy (s) Goal: learn the state values (and maybe the model) In this case: No choice about what actions to take Just execute the policy and learn from experience We’ll get to the general case soon Example: Direct Estimation y Episodes: +100 (1,1) up -1 (1,1) up -1 (1,2) up -1 (1,2) up -1 (1,2) up -1 (1,3) right -1 (1,3) right -1 (2,3) right -1 (2,3) right -1 (3,3) right -1 (3,3) right -1 (3,2) right -1 (3,2) up -1 (4,2) right -100 (3,3) right +100 (done) (done) -100 x = 1, R = -1 U(1,1) ~ (93 + -105) / 2 = -6 U(3,3) ~ (100 + 98 + -101) / 3 = 32.3 Model-Based Learning Idea: Learn the model empirically (rather than values) Solve the MDP as if the learned model were correct Empirical model learning Simplest case: Count outcomes for each s,a Normalize to give estimate of T(s,a,s’) Discover R(s,a,s’) the first time we experience (s,a,s’) More complex learners are possible (e.g. if we know that all squares have related action outcomes, e.g. “stationary noise”) Example: Model-Based Learning y Episodes: +100 (1,1) up -1 (1,1) up -1 (1,2) up -1 (1,2) up -1 (1,2) up -1 (1,3) right -1 (1,3) right -1 (2,3) right -1 (2,3) right -1 (3,3) right -1 (3,3) right -1 (3,2) right -1 (3,2) up -1 (4,2) right -100 (3,3) right +100 (done) (done) -100 x =1 T(<3,3>, right, <4,3>) = 1 / 3 T(<2,3>, right, <3,3>) = 2 / 2 Model-Based Learning In general, want to learn the optimal policy, not evaluate a fixed policy Idea: adaptive dynamic programming Learn an initial model of the environment: Solve for the optimal policy for this model (value or policy iteration) Refine model through experience and repeat Crucial: we have to make sure we actually learn about all of the model