1. Experts Learning and The Minimax Theorem for Zero-Sum Games Maria Florina Balcan December 8th 2011

2. Motivation Many situations involve repeated decision making Deciding how to invest your money (buy or sell stocks) What route to drive to work each day Playing repeatedly a game against an opponent with unknown strategy This course: Learning algos for such settings with connections to game theoretic notions of equilibria

3. Roadmap Last lecture: Online learning; combining expert advice; the Weighted Majority Algorithm. This lecture: Online learning, game theory, minimax optimality.

4. Recap: Online learning, minimizing regret, and combining expert advice. “The weighted majority algorithm”N. Littlestone & M. Warmuth “Online Algorithms in Machine Learning” (survey) A. Blum  Algorithmic Game Theory, Nisan, Roughgarden, Tardos, Vazirani (eds) [Chapters 4]

5. Expert 1 Expert 2 Expert 3 Online learning, minimizing regret, and combining expert advice.

6. Using “expert” advice We solicit n “experts” for their advice. Assume we want to predict the stock market. Can we do nearly as well as best in hindsight? We then want to use their advice somehow to make our prediction. E.g., Note: “expert” ´ someone with an opinion. Will the market go up or down? [Not necessairly someone who knows anything.]

7. Formal model There are n experts. Can we do nearly as well as best in hindsight? Each expert makes a prediction in {0,1} For each round t=1,2, …, T The learner (using experts’ predictions) makes a prediction in {0,1} The learner observes the actual outcome. There is a mistake if the predicted outcome is different form the actual outcome.

8. Weighted Majority Algorithm Instead of crossing off, just lower its weight. – Start with all experts having weight 1. Weighted Majority Algorithm Key Point: A mistake doesn't completely disqualify an expert. –Ifthen predict 1 else predict 0 – Predict based on weighted majority vote.

9. Weighted Majority Algorithm Instead of crossing off, just lower its weight. – Start with all experts having weight 1. Weighted Majority Algorithm Key Point: A mistake doesn't completely disqualify an expert. – Predict based on weighted majority vote. – Penalize mistakes by cutting weight in half.

10. Analysis: do nearly as well as best expert in hindsight If M = # mistakes we've made so far and OPT = # mistakes best expert has made so far, then: Theorem:

11. Randomized Weighted Majority 2.4(OPT + lg n) 2.4(OPT + lg n) not so good if the best expert makes a mistake 20% of the time. Also, generalize ½ to 1- . Can we do better? Equivalent to select an expert with probability proportional with its weight. Yes. Instead of taking majority vote, use weights as probabilities. (e.g., if 70% on up, 30% on down, then pick 70:30) Key Point: smooth out the worst case.

12. Randomized Weighted Majority

13. Formal Guarantee for Randomized Weighted Majority If M = expected # mistakes we've made so far and OPT = # mistakes best expert has made so far, then: Theorem: M ·  OPT + (1/  log(n)

14. Randomized Weighted Majority Solves to:

15. Summarizing E[# mistakes] ·  OPT +  -1 log(n). If set  =(log(n)/OPT) 1/2 to balance the two terms out (or use guess-and-double), get bound of E[mistakes] · OPT+2(OPT ¢ log n) 1/2 Note: Of course we might not know OPT, so if running T time steps, since OPT · T, set ² to get additive loss (2T log n) 1/2 regret E[mistakes] · OPT+2(T ¢ log n) 1/2 So, regret/T ! 0.[no regret algorithm]

16. What if have n options, not n predictors? We’re not combining n experts, we’re choosing one. Can we still do it? Nice feature of RWM: can be applied when experts are n different options We did not see the predictions in order to select an expert (only needed to see their losses to update our weights) E.g., n different ways to drive to work each day, n different ways to invest our money.

17. Decision Theoretic Version; Formal model There are n experts. The guarantee also applies to this model!!! For each round t=1,2, …, T No predictions. The learner produces a prob distr. on experts based on their past performance p t. The learner is given a loss vector l t and incurs expected loss l t ¢ p t. The learner updates the weights. [Interesting for connections between GT and Learning.]

18. Can generalize to losses in [0,1] If expert i has loss l i, do: w i à w i (1-l i  ). [before if an expert had a loss of 1, we multiplied by (1-epsilon), if it had loss of 0 we left it alone, now we do linearly in between] Same analysis as before.

19. “Game Theory, On-line Prediction, and Boosting”, Freund & Schapire, GEB This lecture: Online Learning, Game Theory, and Minimax Optimality

20. Zero Sum Games Game defined by a matrix M. RockPaper 0 1 1/2 Scissors 0 1 101/2 Rock Paper Scissors Row player (Mindy) chooses row i. Column player (Max) chooses column j (simultaneously). Mindy’s goal: minimize her loss M(i,j). Assume wlog entries are in [0,1]. Max’s goal: maximize this loss (zero sum).

21. Randomized Play Mindy chooses a distribution P over rows. Mindy’s expected loss: If i,j = pure strategies, and P,Q = mixed strategies Max chooses a distribution Q over columns [simultaneously] M(P,j) - Mindy’s expected loss when she plays P and Max plays j M(i,Q) - Mindy’s expected loss when she plays i and Max plays Q

22. Sequential Play Say Mindy plays before Max. If Mindy chooses P, then Max will pick Q to maximize M(P,Q), so the loss will be So, Mindy should pick P to minimize L(P). Loss will be: Similarly, if Max plays first, loss will be:

23. Minimax Theorem Playing second cannot be worse than playing first Mindy plays first Von Neumann’s minimax theorem: Mindy plays second No advantage to playing second! Regardless of who goes first the outcome is always the same!

24. Optimal Play Von Neumann’s minimax theorem: 1. Even if Max knows Mindy’s strategy, Max cannot get better outcome than v. v is the best possible value. Optimal strategies: Value of the game Min-max strategy Max-min strategy 9 min-max strategy P * s.t. for any Q, M(P *,Q) · v. 2. No matter what strategy Mindy uses, the outcome is at worst v. 9 max-min strategy Q * s.t. for any P, M(P, Q * ) ¸ v.

25. Optimal Play Von Neumann’s minimax theorem: Optimal strategies: Value of the game Min-max strategy Max-min strategy P * and Q * optimal strategies if the opponent is also optimal! For a two person zero-sum game against a good opponent, your best bet is to find your min-max optimal strategy and always play it.

26. Optimal Play Von Neumann’s minimax theorem: Note: (P *, Q * ) is a Nash equilibrium. Optimal strategies: Value of the game Min-max strategy Max-min strategy All the NE have the same value in zero-sum games. Not true in general, very specific to zero-sum games!!! P * is a best response to Q * ; Q * is a best response to P *

27. Optimal Play Von Neumann’s minimax theorem: Optimal strategies: Value of the game Min-max strategy Max-min strategy P * and Q * optimal strategies if the opponent is also optimal! For a two person zero-sum game against a good opponent, your best bet is to find your min-max optimal strategy and always play it.

28. Beyond the Classic Theory Opponent may not be fully adversarial. M maybe unknown or very large. As game is played over and over, opportunity to learn the game and/or the opponent’s strategy. Often limited info about the game or the opponent Bart Simpson always plays Rock instead of choosing the uniform distribution. You can play Paper and always beat Bart.

29. Repeated Play M unknown. Mindy chooses P t For each round t=1,2, …, T Max chooses Q t (possibly based on P t ) Mindy’s loss is M(P t, Q t ) Mindy observes loss M(i, Q t ) for each pure strategy i. Mindy can run RWM to ensure: where = P t ¢ (MQ t ) l t = MQ t Exactly fits DT experts model! min P M(P,(Q 1 +…+Q T )/T) · v

30. Prove minimax theorem as corollary Imaging game is played repeatedly. In each round t [ ¸ part is trivial ] Define: Need to prove: Mindy plays using RWM Max chooses best response

31. One slide proof of minimax theorem is a strategy that you can use if you have to go first.