1Eric Xing 1Machine LearningMachine Learning1010--701/15701/15--781, Spring 2008781, Spring 2008Reinforcement learning 1Reinforcement learning 1Eric XingEric XingLecture 27, April 28, 2008Reading: Chap. 13, T.M. bookstart3211234+1-1Eric Xing 2Outlinez Intro to reinforcement learningz MDP: Markov decision problemz Dynamic programming:z Value iterationz Policy iteration2Eric Xing 3What is Learning?z Learning takes place as a result of interaction between an agent and the world, the idea behind learning is thatz Percepts received by an agent should be used not only for understanding/interpreting/prediction, as in the machine learning tasks we have addressed so far, but also for acting, and further more for improving the agent’s ability to behave optimally in the future to achieve the goal.Eric Xing 4Types of Learning z Supervised Learningz A situation in which sample (input, output) pairs of the function to be learned can be perceived or are givenz You can think it as if there is a kind teacher- Training data: (X,Y). (features, label)- Predict Y, minimizing some loss.- Regression, Classification.z Unsupervised Learning- Training data: X. (features only)- Find “similar” points in high-dim X-space.- Clustering.3Eric Xing 5Example of Supervised Learning z Predict the price of a stock in 6 months from now, based on economic data. (Regression)z Predict whether a patient, hospitalized due to a heart attack, will have a second heart attack. The prediction is to be based on demographic, diet and clinical measurements for that patient. (Logistic Regression)z Identify the numbers in a handwritten ZIP code, from a digitized image (pixels). (Classification)Eric Xing 6Example of Unsupervised Learningz From the DNA micro-array data, determine which genes are most “similar”in terms of their expression profiles. (Clustering)4Eric Xing 7Types of Learning (Cont’d) z Reinforcement Learningz in the case of the agent acts on its environment, it receives some evaluation of its action (reinforcement), but is not told of which action is the correct one to achieve its goal- Training data: (S, A, R). (State-Action-Reward)- Develop an optimal policy (sequence of decision rules) for the learner so as to maximize its long-term reward. - Robotics, Board game playing programs.Eric Xing 8RL is learning from interaction5Eric Xing 9Examples of Reinforcement Learning z How should a robot behave so as to optimize its “performance”? (Robotics)z How to automate the motion of a helicopter? (Control Theory)z How to make a good chess-playing program? (Artificial Intelligence)Eric Xing 10Robot in a roomz what’s the strategy to achieve max reward?z what if the actions were deterministic?6Eric Xing 11History of Reinforcement Learningz Roots in the psychology of animal learning (Thorndike,1911).z Another independent thread was the problem of optimal control, and its solution using dynamic programming(Bellman, 1957).z Idea of temporal difference learning (on-line method), e.g., playing board games (Samuel, 1959).z A major breakthrough was the discovery of Q-learning (Watkins, 1989).Eric Xing 12What is special about RL?z RL is learning how to map states to actions, so as to maximize a numerical reward over time.z Unlike other forms of learning, it is a multistage decision-making process (often Markovian).z An RL agent must learn by trial-and-error. (Not entirely supervised, but interactive)z Actions may affect not only the immediate reward but also subsequent rewards (Delayed effect).7Eric Xing 13Elements of RLz A policy- A map from state space to action space.- May be stochastic.z A reward function- It maps each state (or, state-action pair) toa real number, called reward. z A value function- Value of a state (or, state-action pair) is thetotal expected reward, starting from that state (or, state-action pair).Eric Xing 14Policy8Eric Xing 15Reward for each step -2Eric Xing 16Reward for each step: -0.19Eric Xing 17Reward for each step: -0.04Eric Xing 18The Precise Goalz To find a policy that maximizes the Value function.z transitions and rewards usually not availablez There are different approaches to achieve this goal in various situations.z Value iteration and Policy iteration are two more classic approaches to this problem. But essentially both are dynamic programming.z Q-learning is a more recent approaches to this problem. Essentially it is a temporal-difference method.10Eric Xing 19Markov Decision ProcessesA Markov decision process is a tuple where:Eric Xing 20The dynamics of an MDPz We start in some state s0, and get to choose some action a0∈Az As a result of our choice, the state of the MDP randomly transitions to some successor state s1, drawn according to s1~Ps0a0z Then, we get to pick another action a1z…11Eric Xing 21The dynamics of an MDP, (Cont’d)z Upon visiting the sequence of states s0, s0, …, with actions a0, a0, …, our total payoff is given byz Or, when we are writing rewards as a function of the states only, this becomesz For most of our development, we will use the simpler state-rewards R(s), though the generalization to state-action rewards R(s; a) offers no special diculties.z Our goal in reinforcement learning is to choose actions over time so as to maximize the expected value of the total payoff:Eric Xing 22Policyz A policy is any function mapping from the states to the actions.z We say that we are executing some policy if, whenever we are in state s, we take action a = π(s).z We also define the value function for a policy πaccording toz Vπ(s) is simply the expected sum of discounted rewards upon starting in state s, and taking actions according to π.12Eric Xing 23Value Functionz Given a fixed policy π, its value function Vπsatisfies the Bellman equations:z Bellman's equations can be used to efficiently solve for Vπ(see later)Immediate rewardexpected sum offuture discounted rewardsEric Xing 24M = 0.8 in direction you want to go0.2 in perpendicular 0.1 left0.1 rightPolicy: mapping from states to actions3211234+1-10.7053211234+1-10.8120.7620.868 0.9120.6600.655 0.611 0.388An optimal policy for the stochastic environment:utilities of states:EnvironmentObservable (accessible): percept identifies the statePartially observableMarkov property: Transition probabilities depend on state only, not on the path to the state.Markov decision problem (MDP).Partially observable MDP (POMDP): percepts does not have enough info to identify transition probabilities.The Grid
View Full Document