You are on page 1of 8

# APPROXIMATION IN ALGORITHMIC GAME THEORY: ROBUST APPROXIMATION BOUNDS FOR EQUILIBRIA AND AUCTIONS

TIM ROUGHGARDEN STANFORD UNIVERSITY SCRIBED BY QIQI YAN

1. Introduction 1.1.

Motivation.

Many modern computer science applications involve autonomous,

self-interested agents. It is therefore important for us to consider agents' strategic behavior in modelling the problems, where non-cooperative game theory can be very helpful. Unfortunately, as one can expect, strategic behavior of the agents often make full optimality dicult or impossible for various reasons. Three common reasons are the following: (1) Equilibria (e.g., Nash) of noncooperative games are typically suboptimal. (2) Auctions lose revenue from strategic behavior. (3) Incentive constraints can make polynomial time approximation of NP-hard problems even harder. As a result, it is essential for approximation, or approximate optimality, to come into play. In particular, approximation is relevant in several topics of study in algorithmic game theory, notably: (1) In the area of price of anarchy, we study worst-case approximation guarantees for equilibria. (2) In revenue maximization, we discuss approximately-optimal revenue guarantees for auctions in non-Bayesian settings. (3) In algorithmic mechanism design, we study approximation algorithms robust to selsh behavior. (4) In computing approximate equilibria, we study the question of whether there is a PTAS for computing an approximate Nash equilibrium. In this scribed talk, we focus on the rst two topics. The third topic was covered in a FOCS 2010 tutorial given by the speaker and available online.

2. Price of Anarchy and Smoothness 2.1.

Price of Anarchy.

The notion of price of anarchy was introduced by Kout-

soupias and Papadimitriou in [26]. It quanties the ineciency of equilibria w.r.t. an objective function. E.g., a typical equilibrium concept is Nash equilibrium, where at equilibrium, no player is better o by switching her strategy unilaterally.

These are scribed notes for a talk given by Tim Roughgarden at an approximation algorithms workshop held at Princeton University in June 2011.
1

and we may not reach equilibrium for the following • • • There may not exist a (pure) equilibrium at all. On the other hand. In Nash equilibrium. 11] The equilibrium might be hard to learn in a distributed way. the reality is is at an equilibrium. some nonnegative minimization objective function) is the supremum of: objective function value at equilibrium optimal objective function value Here the supremum is over all equilibria. The POA for maximizing problems can be similarly dened. 2] correlated equilibria [12] 1I thank Kshipra Bhawalkar for providing the nicely drawn gures. Therefore. then the outcome is near-optimal.1. Each edge in the network has a congestion function. Nash Equilibrium and Optimal A remark here is that there might exist multiple equilibria. with total cost of 24. we look at the worst one w. The price of anarchy (POA) of a game (w. An equilibrium can be computationally hard to nd.r. A price of anarchy bound means that if the game However. 2x s 5 0 12 t 5x s 2x 0 5 12 t 5x Figure 2. where there are two players who 1 want to go from s to t. 2. as a function of the total number of players using that edge. summing up to 28 in total. 15. . we do need to look for more robust guarantees than simple POA bounds to address these concerns. potential reasons: often much more complicated. . Such outcomes could be: • • best-response dynamics. Consider the following network . Either way. pre-convergence outcomes [28. 20. When this is the Call for Robustness.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 2 Denition 1. These raise the worry: are our price of anarchy bounds meaningless? We do not think that these concerns render price of anarchy bounds meaningless.t. while the other player should take the bottom path. [19.t. one player should take the top path.r. However. case.2. Each player chooses exactly one path. a ratio closer to 1 is more desirable. which is the cost for each player using that edge. Our high-level goal is the following: we want worst-case bounds that apply even to non-equilibrium outcomes. our objective. both players choose the same path. to minimize total cost. incurring a cost of 2×2+5×2 each. Example 2. the price of anarchy in this case is 28 24 = 7 6.

λ then the price of anarchy is at most 1−µ . Note that smoothness condition is a stronger condition than is necessary to derive the above bound on the price of anarchy.a. i. 1. s is not a pure Nash For the settings listed below. We assume that a game is ∗ (λ. then the sum of all such costs is upper-bounded jointly using costs of s and s∗ . Let Proof. Such an argument is called a smoothness argument. s−i ) i i ≤ ≤ λ · cost(s∗ ) + µ · cost(s) Here the rst line is by the additive denition of total cost. 14] . 12. 2.e. where n s = (si )i We assume that the objective function is the total cost cost(s) = i Ci (s). i Ci (s). and the third line is by the smoothness cost(s) ≤ λ 1−µ · cost(s∗ ). the following condition holds: Ci (s∗ . We remark that it is important that the cost is additive.k. players. the second line is by the fact that condition. To make sense out of this condition.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 3 • coarse correlated equilibria a. and s be an optimal solution. each picking a strategy is the strategy vector. si . µ < 1) if for every pair s. the price of λ anarchy of pure equilibria is at most 1−µ . 30. Our smoothness condition is the following: Denition 3. Theorem 4. s−i ) ≤ λ · cost(s∗ ) + µ · cost(s) i i This says that if we start with a strategy prole the cost of s. A game is (λ. Remark 5. µ)-smooth (λ > 0. price of anarchy bounds can be obtained via smoothness argumenta in this same canonical way. In the setup. and s∗ is optimal. let If a game is (λ.. price of total anarchy a. µ)-smooth.  no-regret players [7. • • atomic (unweighted) selsh routing [3. s∗ of outcomes. 31] nonatomic selsh routing [32. Then the following holds: cost(s) = i Ci (s) Ci (s∗ . Smoothness and Price of Anarchy.k. µ)-smooth.3. s be a Nash equilibrium. Player To dene our smoothness framework. it follows that s is a Nash equilibrium. the smoothness condition requires the inequality to hold even when the prole equilibrium. Now by rearranging. one only needs the smoothness condition to hold in the special case where s is a Nash equilibrium. let there be incurs a cost of we need to introduce a somewhat abstract setup. and consider for each agent i i if she unilaterally deviates to her strategy in prole s∗ . On the other hand. in a canonical way. 8] A smoothness condition will be the key to our goal.a. To derive the price of anarchy bound. we rst see how the smoothness condition implies price of anarchy bounds.

µ) such that all such games are (λ. in the limit as T goes to innity. µ)-smooth.e. etc.g. . and the proof nishes by rearranging.5. . Another direction is to study POA bounds for auctions. 6] submodular maximization games [33. 25]) from modest intervention (e. st ) + ∆i. 24]. .4.t Here the rst equality is by linear separability of cost. [10. Theorem 7 (An Out-Of-Equilibrium Bound) [31] In a (λ..). [4]). . etc. What's Next. One direction is to beat the worst-case POA bounds by always reaching a nonworst-case equilibrium. 27] coordination mechanisms [13] Smoothness arguments can be used to prove Beyond Nash Equilibria. −i i ∆i. sT denote a no-regret outcome sequence. (sketch) Let Assuming s1 .) Proof. cost(st ) = t we have: Ci (st ) i t = t i [Ci (s∗ . combinatorial auctions. . weighted congestion games [1. s∗ ). It is well-known for example that if every player uses the multiplicative weights algorithm. and therefore it would be nice if we can have bounds on their (Bayes-Nash) equilibria [5. unweighted congestion games C are tight in the following sense: the maximum over congestion games with cost functions in C is equal to the over (λ. Practical auctions often lack "dominant strategies" (like in sponsored search. 2. There are several future research directions along these lines. the same upper bound proved for . and dividing using Furthermore. We illustrate the idea here using the example of no-regret outcomes. the λ average cost of every no-regret sequence is at most 1−µ times the cost of an optimal outcome. s2 . pure Nash equilibria. . i. and let s∗ denote the optimal outcome. then o(1) regret is achieved in polynomial time.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 4 • • • 2. s2 . A sequence worst-case approximation guarantees for outcomes that are not Nash equilibria. . µ)-smooth game. (λ.t ] i −i [λ · cost(s∗ ) + µ · cost(st )] + t i t ≤ ∆i.t ≤ o(T ) T. the connection of the price of anarchy to smoothness is intrinsic: Theorem 8. (I. and the inequality t for each i. Ci (st ) − Ci (s∗ . Now by the no-regret assumption. One can show that weighted congestion games [6] are also tight in this sense. whereas in the second equality. [31] For every set with cost functions restricted to of the pure POA λ minimum of 1−µ C of cost functions. .g. the regret term∆i. µ)-smoothness.t is dened as follows from the smoothness condition. the average cost player i incurs over the sequence is (in the limit as T goes to innity) no worse than playing qi every time. s1 . Non-worst-case equilibria might be reached due to learning dynamics (e. sT of the player of outcomes is no-regret if for each player and each xed action qi i. Denition 6. .

as dened in Myerson [29].1. where we compare revenue to that of an opponent with statistical information about input. and can be extended to approximate equilibria. and Bayes-Nash equilibria. but only used in its analysis. local smoothness for correlated equilibria. One interpretation of the theorem is that a small increase in competition is more important than running the optimal auction.d. The classical Bulow-Klemperer theorem can be seen as one such prior-independent result.d. We have two bidders and one item for sale.2. Myerson characterized the optimal auction for this context. and in the rest of this section we focus on the Bulow-Klemperer theorem. We consider the prior-independent analysis The goal is then to design a distribution- framework of [21. a novel analysis framework is guarantee a meaningful fraction of this benchmark. The classical Bulow-Klemperer theorem [9] says the following: Theorem 9. This also in some sense gives a "bicriteria bound"..englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 2. auction with [9] For every n. smoothness bounds provably give optimal POA bounds in fundamental cases. Note that the Bulow-Klemperer theorem is prior-independent. smoothness is a  canon- ical way to bound the price of anarchy (for pure equilibria).d. 3. 3. in that the Vickrey auction does not need knowledge about the underlying distribution. independent auction that is always near-optimal for the underlying distribution (no matter what the distribution is). Let us conFor sider a very simple problem. which auction (e. i has a valuation vi for winning the item. 22]. needed. best-response dynamics. Note that here distribution over inputs is not used in the design of the auction. . Furthermore. One could try to use competitive analysis: compare your revenue to that obtained by an omniscient opponent.) Let us consider the auction problem of selling a single item to one of Each bidder n bidders. (Technically. it is easy to prove that no auction can Therefore. To summarize this section. from a distribution F. A New Analysis Framework.2.i. Failure of Traditional Competitive Analysis in Auctions. the Bulow-Klemperer theorem is not exactly the same as a prior-independent approximation guarantee as it needs the additional bidder. where the vi 's are drawn i.1. F . But this need can be easily eliminated [17] at a constant loss in the approximation factor. such problems.g. It gives robust POA bounds in the sense that smoothness bounds extend automatically beyond Nash equilibria. 5 Summary of Key Points.i. 3. We start by pointing out the failure of traditional competitive analysis in this context.i. Auctions In this section we study revenue-maximizing auctions. the expected revenue of Vickrey's auction with n+1 i. whether to pick an opening bid of 1 cent or 10 dollars) is best heavily depends on the (unknown) input bids from the bidders. However. 3. and regular F. and all auctions are equally terrible in terms of competitive ratio. bidders from F. Bulow-Klemperer Theorem.6. bidders from F is at least the expected revenue of the optimal n i. where the optimal auction is dependent on the distribution We make the standard regularity condition on the distribution F.

) Recall that the Bulow-Klemperer theorem says that for every pected revenue over revenue from of p∗ . It turns out that the revenue function R(q) is concave in q if and only if the distribution F is regular.i. or the height of the curve. the revenue of using a random reserve drawn from F equals to the expected value of R(q) with q uniformly drawn from [0. Vickrey's exn+1 i. the F −1 (1 − q). bidders is as good as the maximum-possible expected n i. and the expected revenue is R(q).g. she faces a take-it-or-leave-it oer which equals to the bid of the other bidder. 1].2. I. the area under the curve is at least half of the height of the curve. we rst need to understand the optimal auction given distributional information. We can solve for the optimal p∗ . twice the expected revenue from one bidder with a random reserve (drawn from reserve F ) is as good as the expected revenue from one bidder with an optimal One advantage of this reformulation is that the problem is now purely about selling one item to one bidder. 2 But what about cases beyond one bidder and one item (and i. ∗ ∗ Now let q maximize R(q). bidders. n = 1 can be reformulated from each bidder's perspective. The value of the bidder for the item is drawn from a known distribution price v F. which is in some sense the most interesting and non-trivial case. (Note: n. the expected revenue of p is p(1 − F (p)).i.2. which is in turn is a random draw from distribution Therefore. the Bulow-Klemperer theorem for the case of as follows: Claim: r∗ . In the following. Suppose we oer price p. Revenue Curve .d. Let n = 1 for now. consider one bidder and one item. if we sell at a price so that the probability of sale is q .d. First.. completing the Figure 3.e. To prove the Bulow-Klemperer theorem. We dene the (expected) revenue function price has to be R(q) = q · F −1 (1 − q) as a function of sale probability.1. Then the optimal revenue is R(q ). [29] The auction that gives the maximum expected revenue is the second-price auction (or Vickrey auction) with the above-dened reserve p∗ is independent of n. p∗ = 1 for v ∼ unif orm[0. and we want to choose an optimal reserve p to maximize revenue. bidders. q axis.i.i. On the other hand. F. which equals to the area under the curve and above the proof. 1]. Note that in a Vickrey auction with n + 1 = 2 i.d. e. we prove this theorem for the case n = 1.d. By concavity of R(q). Bayesian Prot Maximization. draws vi )? Myerson characterized the optimal auction as follows: Theorem 10.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 6 3.

Saks. Weighted congestion games: Price of anarchy. J. SIAM. In Mark de Berg and Ulrich Meyer. Skopalik. 2010. and Shang-Hua Teng.d. MohammadTaghi Hajiaghayi. Online multicast with egalitarian cost sharing. Martin Gairing. But we can get a simultaneous competitive guarantee with all Bayesian-optimal auctions. and A. and Aaron Roth.g. D. ACM. Welfare guarantees for combinatorial auctions with item bidding. 2011. 2005. 86(1):180194. Auctions versus negotiations. . Schoppmann. to economics and game theory). [6] Kshipra Bhawalkar. Tsinghua University Press. 2008. Charikar. ICS. Mirrkoni. Bhawalkar and T. and studying approximation in various contexts gives not only robust guarantees. pages 218229. settings? Or what about multi-parameter settings. Eyal Even-Dar. 2010. J. M. Azar. V. 23]. 2006. Azar. Avrim Blum. Naor. 7 Related Work. Theory of Computing. universal worst-case examples. and F. In Proceedings of the 23rd Annual Symposium on Theoretical Aspects of Computer Science (STACS). In Proceedings of the 20th ACM Symposium on Parallelism in Algorithms and Architextures (SPAA). 2010. and what about non-i. More general forms of such "bicriteria bounds" were obtained in [22. Bulow and P. Awerbuch. editor. and M. Katrina Ligett. 2008. and Yishay Mansour. and L. Gairing. 3. include "non-downward-closed" ones. and tightness. 17. [10] M. pages 200213. [3] B. pages 373382. but also new insights into the problem. American Economic Review. Monien. The price of routing unsplittable ow. ACM. [11] Xi Chen. Epstein. Standard competitive analysis is useless for worst-case revenue maximization. Roughgarden. and more prior-independent approximations have been achieved in [16. Klemperer. C.i. 2009. like combinatorial auctions? Epilogue The higher-level moral of this survey talk is the following. ESA (2). H.. In Proceedings of the 9th ACM Conference on Electronic Commerce (EC). [4] Maria-Florina Balcan. Worst-case approximation guarantees can be a powerful "intellectual export" from theoretical computer science to other elds (e. There are many reasons why we should consider approximation (not just computational tractability).3. editor. STOC. 56(3). Regret minimization and the price of total anarchy. Settling the complexity of computing twoplayer nash equilibria. Dumrauf. pages 7076. Fast convergence to nearly optimal solutions in potential games. 18]. and Katrina Ligett.4. Y. 1996. editor. pages 264273. SODA. Epstein. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing (STOC). A. Y. editors. In Andrew Chi-Chih Yao. In Cynthia Dwork. Mathieu. can we thoroughly understand problems in the socalled single-parameter domain. There has been a lot of recent progress on prior-independent approximations. Summary. [7] Avrim Blum. In Dana Randall. volume 6347 of Lecture Notes in Computer Science. [8] Avrim Blum. Xiaotie Deng. Circumventing the price of anarchy: Leading dynamics to good behavior. pages 1728. B. Karlo. Toward future directions. pages 5766. [9] J. [2] B. [5] K. Awerbuch. 6(1):179199. volume 3884 of Lecture Notes in Computer Science. and Tim Roughgarden.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 3. 2008. Aland. Routing without regret: On convergence to nash equilibria of regret-minimizing algorithms in routing games. Exact price of anarchy for polynomial congestion games. S. Springer. pages 700709. References [1] S.

How bad is selsh routing? Journal of the ACM. 64(2):457469. and M. Non-price equilibria in markets of discrete goods. 2011. and prot. [18] S. Mirrokni. Mirrokni and A. Goemans. 2008. Stier Moses. In Proceedings of the 43rd Annual Symposium on Foundations of Computer Science (FOCS). Tardos. In ACM Conference on Electronic Commerce. 2005. and É. D. Hartline and T. [14] J. Intrinsic robustness of the price of anarchy. A. Roughgarden and É. on Electronic Commerce (EC). [23] J. In Proceedings of the 10th Conference on Integer Programming and Combinatorial Optimization (IPCO). pages 404413. pages 6773. pages 604612. T. 2009. The complexity of computing a nash equilibrium. 6(1):5873. 2010. and Noam Nisan. Myerson. Yan. Dughmi. Marden and T. In Proc. Envy. ACM. Sink equilibria and convergence. Inner product spaces for minsum coordination mechanisms. [32] T. 2004. USA. R. Vetta. D. pages 295296. [27] J. Paul W. V. pages 7584. pages 225234. Gkatzelis. New York. Convergence issues in competitive games. Optimal mechanism design and money burning. Yan. 39(1):195259. [24] Avinatan Hassidim. Papadimitriou. and Q. 1981. Yishay Mansour. G. Revenue maximization with a single sample. Goldberg. pages 513522. Generalized eciency bounds in distributed resource allocation. on Electronic Commerce (EC). The complexity of pure Nash equilibria. On the ineciency of equilibria in congestion games. Piliouras. SIAM J.englishAPPROXIMATION IN ALGORITHMIC GAME THEORY 8 [12] G. pages 4150. In Proceedings of the 36th Annual ACM Symposium on Theory of Computing (STOC). Sundararajan. In Proceedings of the 41st Annual ACM Symposium on Theory of Computing (STOC). volume 3064 of Lecture Notes in Computer Science. In Proc. Roughgarden. Koutsoupias. In Proceedings of the Conference on Decision and Control (CDC). In STOC. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing (STOC). [30] G. Hartline and T. Cole. R. Schulz. In Proc. 2002. on Theory of Computing (STOC). [33] A. 2008. Correa. C. NY. and K. 2009. Games and Economic Behavior. volume 1563 of Lecture Notes in Computer Science. 39th ACM Symp. T. Roughgarden. . Dhangwatnotai. Hartline and Q. Mirrokni. and N. IEEE. and N. truth. H. 49(2):236 259. 2004. In STOC. on Electronic Commerce (EC). pages 539548. In Proc. In Proceedings of the 46th Annual Symposium on Foundations of Computer Science (FOCS). 12th ACM Conf. [15] Constantinos Daskalakis. In Proceedings of the 16th Annual Symposium on Theoretical Aspects of Computer Science (STACS). [28] V. [31] T. 1999. D. [29] R. Revenue submodularity. D. Roughgarden. 2011. Comput. R. [16] N. Christodoulou and E. and A. 2009. Roughgarden. 2010. S. 2009. 2009. [17] P. pages 2233 2238. Limited and online supply and the bayesian foundations of prior-free mechanism design. Perakis. with applications to facility location. V. Tardos. pages 243252. 2004. X. Olver. E. Worst-case equilibria. Vetta. The price of anarchy when costs are non-separable and asymmetric. pages 4658. [13] R. The price of anarchy of nite congestion games. 2011. Papadimitriou. Fabrikant. trafc routing and auctions. Koutsoupias and C. Nash equilibria in competitive societies. S. [21] J. Roughgarden. V. Talwar. 11th ACM Conf. [19] A. 11th ACM Conf. H. In EC '09: Proceedings of the tenth ACM conference on Electronic commerce. J. In ACM Conference on Electronic Commerce. pages 533542. Correa. Hartline. pages 243252. [22] J. 2005. pages 129138. [26] E. 2009. Papadimitriou. Devanur and J. Simple versus optimal mechanisms. Roughgarden. Haim Kaplan. Kleinberg. Mathematics of Operations Research. Vetta. D. In Proceedings of the 7th International Workshop on Approximation Algorithms for Combinatorial Optimization Problems (APPROX). Optimal auction design. and Christos H. Multiplicative updates outperform generic noregret learning in congestion games. S. pages 416425. 2002. [20] M. pages 142151. [25] R. pages 183194.