1. CS 541: Artificial Intelligence Lecture XI: Reinforcement Learning Slides Credit: Peter Norvig and Sebastian Thrun

2. Schedule  TA: Vishakha Sharma  Email Address: vsharma1@stevens.edu WeekDateTopicReadingHomework 108/30/2011Introduction & Intelligent AgentCh 1 & 2N/A 209/06/2011Search: search strategy and heuristic searchCh 3 & 4sHW1 (Search) 309/13/2011Search: Constraint Satisfaction & Adversarial SearchCh 4s & 5 & 6 Teaming Due 409/20/2011Logic: Logic Agent & First Order LogicCh 7 & 8sHW1 due, Midterm Project (Game) 509/27/2011Logic: Inference on First Order LogicCh 8s & 9 610/04/2011Uncertainty and Bayesian NetworkCh 13 &14sHW2 (Logic) 10/11/2011No class (function as Monday) 710/18/2011Midterm Presentation Midterm Project Due 810/25/2011Inference in Baysian NetworkCh 14sHW2 Due, HW3 (Probabilistic Reasoning) 911/01/2011Probabilistic Reasoning over TimeCh 15 1011/08/2011TA Session: Review of HW1 & 2 HW3 due, HW4 (MDP) 1111/15/2011Supervised Machine LearningCh 18 1211/22/2011Markov Decision ProcessCh 16 1311/29/2011Reinforcement learningCh 21HW4 due 1412/06/2011Final Project Presentation Final Project Due

3. Recap of Lecture X  Planning under uncertainty  Planning with Markov Decision Processes  Utilities and Value Iteration  Partially Observable Markov Decision Processes

4. Review of MDPs

5. Review: Markov Decision Processes  An MDP is defined by:  set of states s  S  set of actions a  A  transition function P(s’| s, a) aka T(s,a,s’)  reward function R(s, a, s’) or R(s, a) or R(s’)  Problem:  Find policy π(s) to maximize  ∑ t R(s, π(s), s′ ) 5

6. Value Iteration  Idea:  Start with U 0 (s) = 0  Given U i (s), calculate:  Repeat until convergence  Theorem: will converge to unique optimal utilities  But: Requires model (R and P) 6

7. What is reinforcement learning

8. Reinforcement Learning  What if the state transition function P and the reward function R are unknown? 8 http://heli.stanford.edu/

9. Example: Backgammon  Reward only for win / loss in terminal states, zero otherwise  TD-Gammon (Tesauro, 1992) learns a function approximation to U(s) using a neural network  Combined with depth 3 search, one of the top 3 human players in the world – good at positional judgment, but mistakes in end game beyond depth 3 9

10. 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 10

11. What to Learn  If you do not know P, R, what you learn depends on what you want to do  Utility-based agent: learn U(s)  Q-learning agent: learn utility Q(s, a)  Reflex agent: learn policy π (s)  And on how patient/reckless you are  Passive: use a fixed policy π  Active: modify policy as you go 11

12. Passive Temporal-Differences

13. Passive Temporal-Difference 13

14. Example 14 +1 000 00 000 0

15. Example 15 +1 000 00 000 0

16. Example 16 +1 000 00 0 0.30 0 0

17. Example 17 +1 000 00 0 0.30 0 0

18. Example 18 +1 000 00 0.0.0675 0.30 0 0

19. Example 19 +1 -0.01-.16.12 -.12.17.28.36.20 -.2

20. Sample results 20

21. Problems in passive learning

22. Problem  The problem  Algorithm computes the value of one specific policy  May never try the best moves  May never visit some states  May waste time computing unneeded utilities 22

23. Quiz  Is there a fix?  Can we explore different policies (active reinforcement learning), yet still learn the optimal policy? 23

24. Active reinforcement learning

25. Greedy TD-Agent  Learn U with TD-Learning  After each change to U, solve MDP to compute new optimal policy π (s)  Continue learning with new policy 25

26. Greedy Agent Results  Wrong policy  Too much exploitation, not enough exploration  Fix: higher U(s) for unexplored states 26

27. Exploratory Agent  Converges to correct policy 27

28. Alternative: Q-Learning  Instead of learning U(s), learn Q(a,s): the utility of action a in state s  With utilities, need a model, P, to act:  With Q-values, no model needed  Compute Q-values with TD/exploration algorithm 28

29. Q-Learning 29 000+1 000 000 00 00 00 00 00 00 000 00 000 0 0 0 0 0 0 0

30. Q-Learning 30 000+1 000 000 00 00 00 00 00 00 000 00 000 0 0 0 0 0 0 0

31. Q-Learning 31 000+1 000 000 00 00 00 00 00 00 00?? 00 000 0 0 0 0 0 0 0

32. Q-Learning 000+1 000 000 00 00 00 00 00 00 00.45 00 000 0 0 0 0 0 0 0

33. Q-Learning 000+1 000 000 00 00 00 00 00 00 0.33.78 00 000 0 0 0 0 0 0 0

34. Q-Learning  Q-learning produces tables of q-values: 34

35. Q-Learning Properties  Amazing result: Q-learning converges to optimal policy  If you explore enough  If you make the learning rate small enough  … but not decrease it too quickly!  Basically doesn’t matter how you select actions (!)  Neat property: off-policy learning  learn optimal policy without following it (some caveats, see SARSA) SE SE 35

36. Practical Q-Learning  In realistic situations, we cannot possibly learn about every single state!  Too many states to visit them all in training  Too many states to hold the q-tables in memory  Instead, we want to generalize:  Learn about some small number of training states from experience  Generalize that experience to new, similar states  This is a fundamental idea in machine learning, and we’ll see it over and over again 36

37. Generalization

38. Generalization Example  Let’s say we discover through experience that this state is bad:  In naïve Q learning, we know nothing about this state or its Q states:  Or even this one! 38

39. Feature-Based Representations  Solution: describe a state using a vector of features  Features are functions from states to real numbers (often 0/1) that capture important properties of the state  Example features:  Distance to closest ghost  Distance to closest dot  Number of ghosts  1 / (dist to dot) 2  Is Pacman in a tunnel? (0/1)  …… etc.  Can also describe a Q-state (s, a) with features (e.g. action moves closer to food) 39

40. Linear Feature Functions  Using a feature representation, we can write a Q function (or Utility function) for any state using a few weights:  Advantage: our experience is summed up in a few powerful numbers  Disadvantage: states may share features but be very different in value! 40

41. Function Approximation  Q-learning with linear Q-functions:  Intuitive interpretation:  Adjust weights of active features  E.g. if something unexpectedly bad happens, disprefer all states with that state’s features (prudence or superstition?) 41

42. Summary: MDPs and RL  If we know the MDP  Compute U*, Q*,  * exactly  Evaluate a fixed policy   If we don’t know the MDP  We can estimate the MDP then solve  We can estimate U for a fixed policy   We can estimate Q*(s,a) for the optimal policy while executing an exploration policy 42  Model-based DP  Value Iteration  Policy iteration  Model-based RL  Model-free RL:  Value learning  Q-learning Things we know how to do: Techniques: