1. Guiding dynamics in potential games Avrim Blum Carnegie Mellon University Joint work with Maria-Florina Balcan and Yishay Mansour [Cornell CSECON 2009] [This talk based on results in “Improved Equilibria via Public Service Advertising”, SODA’09 and “The Price of Uncertainty”, ACM-EC’09]

2. Good equilibria, Bad equilibria Many games have both good and bad equilibria. In some places, everyone throws their trash on the street. In some, everyone puts their trash in the trash can. In some places, everyone drives their own car. In some, everybody uses and pays for good public transit.

3. Good equilibria, Bad equilibria Many games have both good and bad equilibria. A nice formal example is fair cost-sharing. n players in weighted directed graph G. Player i wants to get from s i to t i, and they share cost of edges they use with others. s t 1n Good equilibrium: all use edge of cost 1. (cost 1/n per player) Bad equilibrium: all use edge of cost n. (cost 1 per player)

4. Good equilibria, Bad equilibria Many games have both good and bad equilibria. A nice formal example is fair cost-sharing. n players in weighted directed graph G. Player i wants to get from s i to t i, and they share cost of edges they use with others. … 1 111 s1s1 snsn t 000 k ¿ n cars Shared transit

5. High-level questions 1. Can a helpful authority encourage behavior to move from bad to good? –Model as having some limited powers of persuasion … 1 111 s1s1 snsn t 000 k ¿ n cars Shared transit

6. High-level questions 1. Can a helpful authority encourage behavior to move from bad to good? –Model as having some limited powers of persuasion 2. In reverse direction, if we get people into a good equilibrium (and players are selfish, reasonably myopic, etc) then like to think behavior will stay there.

7. High-level questions 1. Can a helpful authority encourage behavior to move from bad to good? –Model as having some limited powers of persuasion 2. If game has small fluctuations in costs, or a few Byzantine players, (when) could behavior spiral out of control?

8. Direction 1: guiding from bad to good “Public service advertising model”: 0. n players begin in some arbitrary configuration. 1.Authority launches public-service advertising campaign, proposing joint action s *. … 1 111 s1s1 snsn t 000 k k

9. Direction 1: guiding from bad to good “Public service advertising model”: 0. n players begin in some arbitrary configuration. 1.Authority launches public-service advertising campaign, proposing joint action s *. Each player i pays attention and follows with probability . Call these the receptive players … 1 111 s1s1 snsn t 000 k 1.Authority launches public-service advertising campaign, proposing joint action s *.

10. Direction 1: guiding from bad to good “Public service advertising model”: 1.Authority launches public-service advertising campaign, proposing joint action s *. Each player i pays attention and follows with probability . Call these the receptive players 2.Remaining (non-receptive) players fall to some arbitrary equilibrium for themselves, given play of receptive players. 3.Campaign wears off. Entire set of players follows best- response dynamics from then on. 0. n players begin in some arbitrary configuration. … 1 111 s1s1 snsn t 000 k

11. Direction 1: guiding from bad to good “Public service advertising model”: 1.Authority launches public-service advertising campaign, proposing joint action s *. Each player i pays attention and follows with probability . Call these the receptive players 2.Remaining (non-receptive) players fall to some arbitrary equilibrium for themselves, given play of receptive players. 3.Campaign wears off. Entire set of players follows best- response dynamics from then on. 0. n players begin in some arbitrary configuration. Note #1: if  =1, can just propose best Nash equilibrium. Key issue: what if  < 1?

12. Direction 1: guiding from bad to good “Public service advertising model”: 1.Authority launches public-service advertising campaign, proposing joint action s *. Each player i pays attention and follows with probability . Call these the receptive players 2.Remaining (non-receptive) players fall to some arbitrary equilibrium for themselves, given play of receptive players. 3.Campaign wears off. Entire set of players follows best- response dynamics from then on. 0. n players begin in some arbitrary configuration. Note #2: Can replace 2 with poly(n) steps of best-response for non-receptive players.

13. Direction 1: guiding from bad to good Price of stability: how much worse than optimal can the best Nash equilibrium be? Price of anarchy: how much worse than optimal can the worst Nash equilibrium be? E.g., fair cost sharing: may exist NE as bad as n times worse than OPT, but always exists one at most O(log n) times worse than OPT.

14. Main Results If only a constant fraction  of the players follow the advice, then we can still get within O(log(n)/  ) of OPT. Extend to cost-sharing + linear delays. For any  < 1, an  fraction is not sufficient. Ratio to OPT can still be unbounded. (PoS = log(n), PoA = n) (PoS = 1, PoA = 1 ) (PoS = 1, PoA =  (n 2 )) Threshold behavior: for  > ½, can get ratio O(1), but for  < ½, ratio stays  (n 2 ). (assume degrees  (log n)).

15. Fair Cost Sharing If only an  probability of players following the advice, then we get within O(log(n)/  ) of OPT. (PoS = log(n), PoA = n) - In any NE for non-receptive players, any such player i can’t improve by switching to his path P i OPT in OPT. - Advertiser proposes OPT (any apx also works) #receptives on edge e - Calculate total cost of these guaranteed options. Rearrange sum... … 1 111 s1s1 snsn t 000 k

16. Fair Cost Sharing If only an  probability of players following the advice, then we get within O(log(n)/  ) of OPT. (PoS = log(n), PoA = n) - Calculate total cost of these guaranteed options. Rearrange sum... - Finally, use: X ~ Bi(n,p) - Take expectation, add back in cost of receptives: get O(OPT/  ). (End of phase 2)

17. Fair Cost Sharing If only an  probability of players following the advice, then we get within O(log(n)/  ) of OPT. (PoS = log(n), PoA = n) - Finally, use: X ~ Bi(n,p) - Take expectation, add back in cost of receptives: get O(OPT/  ). (End of phase 2) - Finally, in last phase, std potential argument shows behavior cannot get worse by more than an additional log(n) factor. (End of phase 3)

18. Cost Sharing, Extension + linear delays: - Problem: can’t argue as if remaining NR players didn’t exist since they add to delays - Define shadow game: pure linear latency fns. Offset defined by equilib at end of phase 2. # users on e at end of phase 2 - Behavior at end of phase 2 is equilib for this game too. - Show - This has good PoA.

19. Party affiliation games Given graph G, each edge labeled + or -. Vertices have two actions: RED or BLUE. Pay 1 for each + edge with endpoint of different color, and each – edge with endpoint of same color. Special cases: + + + -- All + edges is consensus game. All – edges is cut-game. +1 to keep ratios finite

20. Party affiliation games (PoS = 1, PoA =  (n 2 )) - Threshold behavior: for  > ½, can get ratio O(1), but for  < ½, ratio stays  (n 2 ). (assume degrees  (log n)). - Consensus game, two cliques, with relatively sparse between them. Players “locked” into place. Lower bound: Degree (1/2 -  )n/8 across cut

21. Party affiliation games (PoS = 1, PoA =  (n 2 )) - Threshold behavior: for  > ½, can get ratio O(1), but for  < ½, ratio stays  (n 2 ). (assume degrees  (log n)). - Split nodes into those incurring low-cost vs those incurring high-cost under OPT. - Show that low-cost will switch to behavior in OPT. For high-cost, don’t care. - Cost only improves in final best-response process. Upper bound:

22. High-level questions 1. Can a helpful authority encourage behavior to move from bad to good? –Model as having some limited powers of persuasion 2. If game has small fluctuations in costs, or a few byzantine players, could behavior spiral out of control?

23. Direction #2 A few ways this could happen: –Small changes cause good equilibria to disappear, only bad ones left. (economy?) –Bad behavior by a few players causes pain for all (nukes) –Neither of above, but instead through more subtle interaction with dynamics… If game has small fluctuations in costs, or a few Byzantine players, could behavior spiral out of control?

24. Model Players follow best (or better) response dynamics. Costs of resources can fluctuate between moves: c i t 2 [c i /(1+  ), c i (1+  )] (alternatively, one or more Byzantine players who move between time steps) Play begins in a low-cost state. How bad can things get? Price-of-Uncertainty (  ) of game = maximum ratio of eventual social cost to initial cost.

25. Model Players follow best (or better) response dynamics. Costs of resources can fluctuate between moves: c i t 2 [c i /(1+  ), c i (1+  )] Price-of-Uncertainty (  ) of game = maximum ratio of eventual social cost to initial cost.

26. Model Players follow best (or better) response dynamics. Costs of resources can fluctuate between moves: c i t 2 [c i /(1+  ), c i (1+  )] Price-of-Uncertainty (  ) of game = maximum ratio of eventual social cost to initial cost. One way to look at this: Define graph: one node for each state. Edge u ! v if perturbation can cause BR to move from u to v. What do the reachable sets look like?

27. Set-cover games Special case of fair cost-sharing n players, m resources, with costs c 1,…,c m. Each player has some allowable resources Each player chooses some allowable resource. Players split cost with all others choosing same one. c1c1 c2c2 c3c3 cmcm

28. Main results Set-cover games: If  = O(1/nm) then PoU = O(log n). However, for any constant  > 0, PoU =  (n). Also, a single Byzantine player can take state from a PNE of cost O(OPT) to one of cost  (n ¢ OPT).

29. Main results General fair-cost-sharing games: If many players for each (s i,t i ) pair (n i =  (m)), then PoU = O(1) even for constant  >0. Open for general number of players. Matroid congestion games: Matroid congestion games: (strategy sets are bases of matroid. E.g., set-cover where choose k resources) If  = O(1/nm) then PoU = O(log n) for fair cost- sharing. In general, if  = O(1/nm) then PoU = O(GAP). In both cases, require best-response. Better- response not enough. (unlike set-cover) Also results for other classes of games too.

30. Set-Cover games (upper bound) For upper bound, think of players in sets as a stack of chips. View i th position in stack j as having cost c j /i. Load chips with value equal to initial cost. When player moves from j to k, move top chip. Cost of position goes up by at most (1+  ) 2. cjcj ckck At most mn different positions. So, following the path of any chip and removing loops, cost of final set is at most (1+  ) 2nm times its value. So, if  = O(1/nm) then PoU = O(log n).

31. Matroid games In matroid games, can think of each player as controlling a set of chips. Nice property of best response in matroids: –Can always order the move so that each individual chip is doing better-response. Apply previous argument. Fails for better-response. –Here, can get player to do kind of binary counting, bad even for exponentially-small .

32. Open questions and directions Looking at: how can we help players find their way to a good state? Getting to good states: nice line of work on how players might be able to do it all by themselves. [Blume, Young, Shamma, Marden, Beggs…] Noisy best-response / noisy adaptive play. Distribution in limit favors good states, like simulated annealing. But, time could be exponential (subway). And how dangerous could small fluctuations be in knocking them out?

33. Open questions and directions Getting to good states: nice line of work on how players might be able to do it all by themselves. [Blume, Young, Shamma, Marden, Beggs…] Noisy best-response / noisy adaptive play. Distribution in limit favors good states, like simulated annealing. But, time could be exponential (subway). To reach good states quickly, need to give players more information about game they are playing. More general, self-interested models for this?