FRIEDMAN, E., AND SHENKER, S. Learning and implementation on the internet. In Manuscript. New Brunswick: Rutgers University, Department of Economics (1997).  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... games [21, 25] potential games [17] and summarization games [14] In addition, several recent papers have proposed new game theoretic ideas to better model distributed and asynchronous networks in which players have little information about their strategic environment, such as the Internet [10, 18, 27]. For further discussion and references on these topics, see [24] We make no attempt at a detailed survey of previous work on the optimal pricing of resources, and instead refer the reader to [6, 13] for recent collections of surveys on pricing by the transportation Problem Studied Linear ....

E. J. Friedman and S. Shenker. Learning and implementation on the Internet. Working paper, 1997.

....used in the mechanism design literature; there are many more solution concepts that we have not described. The question of which solution concept is most appropriate for Internet mechanisms has been debated by many researchers (see, e.g. NR01] and was studied at length by Friedman and Shenker [FS97] Strategyproofness seems to be a safe choice, provided collusion between agents is impractical, because the incentive compatibility of the mechanism then holds regardless of the extent of any agent s knowledge of other agents preferences and strategies. Group strategyproofness will give ....

Eric Friedman and Scott Shenker. Learning and implementation in the internet. Preprint. Available at http://www.icir.org/shenker/decent.ps, 1997.

.... games [22, 26] potential games [18] and summarization games [14] In addition, several recent papers have proposed new game theoretic ideas to better model distributed and asynchronous networks in which players have little information about their strategic environment, such as the Internet [10, 19, 28]. For further discussion and references on these topics, see [25] We make no attempt at a detailed survey of previous work on the optimal pricing of resources, and instead refer the reader to [6, 13] for recent collections of surveys on pricing by the transportation science literature, and to ....

E. J. Friedman and S. Shenker. Learning and implementation on the Internet. Working paper, 1997.

....special class of externality games is considered: Definition 3.1. A non atomic game is a game with a continuum of users, in which no single player can affect the other players by changing only her own action. A nonatomic externality game (NAEG) is a nonatomic game with submodular payoff functions [3]. The key restriction on a representative user i s payoff u i is that it only depends on her own strategy y i (resource usage) and the externality vector fy j g j 6=i : u i (y i ; f(y) a i y i f(y) a i y i P j y j (13) Here by definition u i is a submodular function i.e. function ....

Friedman, E. and Shenker, S., "Learning and implementation on the Internet", Mimeo, 1998

....et al. (1994) Denote now by Opt(r) the function which, for any given reservation price, returns the optimal bid (in a stochastically stable world) Let L be a learning algorithm, where L(r t , h t 1 ) returns period t s bid given a history. Following (but slightly abusing) the terminology of [2] Friedman and Shenker (1998) we can now introduce the following definitions: Optimality: Suppose that the world is stochastically stable. Then we say that a learning algorithm L is optimal if and only if lim t L(r t , h t 1 ) Opt(r t ) 0. Responsiveness: Suppose that a stable ....

Friedman E. J. and Shenker S. (1998), "Learning and Implementation on the Internet", mimeo, Rutgers University.

....C91, D83 # Corresponding author. 1 1 Introduction Limited information environments are found in many economic settings, for example, in distributed networks. In a distributed system such as the Internet agents havevery limited a priori information about other agents and the payo# matrix #Friedman and Shenker 1998#. To design mechanisms for resource allocation in distributed systems it is important to study how agents learn in settings which parallel distributed systems, i.e. under extremely limited information. From the perspective of studying human learning, limited information environments provide the ....

....one output technology with decreasing returns. Eachofthen agents announces his demand q i of output. Let q 1 # q 2 # ### # q n . The cost function is denoted by C, which is strictly convex. The cost sharing mechanism 1 This is a stronger solution concept than dominance solvability. See Friedman and Shenker #1998# for a precise de#nition. 3 must allocate the total cost C# P i q i # among the n agents. Under the serial mechanism #hereafter shortened as SER# agent 1 #with the lowest demand# pays #1=n#th of the cost of producing nq 1 , C#nq 1 #=n. Agent2pays agent 1 s cost share plus 1=#n , 1#th of the ....

Friedman, Eric and Scott Shenker. #Learning and Implementation on the Internet." Manuscript. New Brunswick: Rutgers University, Department of Economics, 1998.

....took a game theoretic approach. Research in this area includes Internet pricing ( 100, 80, 111] and several papers in [40] QoS and resource allocation [20, 93, 51, 87] and network control [45, 38, 12, 11, 7, 44, 101] A game theoretic approach to the behavior of network users can be found in [21, 58]. A recent paper [19] studies a networking problem in the context of mechanism design. Electronic commerce The rapid growth of electronic commerce in recent years drew the attention of several researchers and institutions. Obviously many questions that stem from this area should take into ....

Eric Friedman and Scott Shenker. Learning and implementation on the internet. To appear.

....no matter how other agents behave whether selfish, spiteful, or stupid truthful revelation is the optimal (dominant) strategy for each user. Strategyproofness is much more exacting than Nash implementation, and so Nash implementation can achieve a much wider variety of outcomes. However, [7] argues that strategyproofness may be the only viable approach in the Internet, because one cannot ensure that simultaneous selfish play will reach the traditionally defined notions of equilibrium. There is a long history of applying game theoretic techniques to networking problems. Some use ....

Friedman, E., and S. Shenker (1997). "Learning and Implementation in the Internet," preprint. Available at http://www.aciri.org/shenker/decent.ps

.... [17] Moreover, the set of applications is rapidly expanding, as game like scenarios are emerging in the telecommunications infrastructure, where network links and queues serve as shared resources (see, for example, 19, 26] as well as web sites and other shared databases (see, for example, [13]) In contrast with the recent trend towards proposing solutions to distributed resource allocation problems by pricing congestible resources [20, 21] Arthur suggests that bounded rationality together with learning can yield solutions to problems of resource allocation in decentralized ....

E. Friedman and S. Shenker. Learning and implementation on the Internet. Mimeo, 1997.

....Tauman [1988] for a survey) Most of the theoretical literature have focused on the axiomatic characterization of these mechanisms (see, e.g. Moulin and Shenker [1994] Friedman and Moulin [1995] and their static properties in a complete information setting with synchronous actions. However, as Friedman and Shenker [1997] pointed out, in a distributed system such as the internet where agents have very limited a priori information about other agents and the payoff matrix and where there is no synchronization of actions, traditional solution concepts that we use to characterize these mechanisms, such as Nash ....

....question is the performance of the two mechanisms under incomplete information. This is an important context because of the increasingly important role of the internet in the economy. Exploring feasible mechanisms for allocating costs and resources in the internet has many practical implications. Friedman and Shenker [1997] addresses the issue of learning and implementation on the internet. They argue that when agents have very limited a priori information about the other players and the payoff matrix, standard solution concepts like the Nash equilibrium or even the serially undominated set are not necessarily ....

[Article contains additional citation context not shown here]

Friedman, Eric and Scott Shenker. "Learning and Implementation on the Internet." Manuscript. New Brunswick: Rutgers University, Department of Economics, 1997.

....the computer science literature traditionally did the opposite. The emergence of the Internet as a standard platform for distributed computation has changed this state of a#airs. Incentives have become an increasingly important consideration in network protocol design (see, for example, [FNY89, FS97, HA88, KLO95, KS89, S88, S90, S95]) More recently, the work of Nisan and Ronen [NR01] has inspired the design of algorithmic mechanisms for, e.g. scheduling, load balancing, lowest cost paths, and combinatorial auctions that satisfy both the traditional economic definitions of incentive compatibility and the traditional TCS ....

Friedman, E., and Shenker, S. (1997). "Learning and Implementation in the Internet," preprint. Available at http://www.icir.org/shenker/decent.ps

....the computer science literature traditionally did the opposite. The emergence of the Internet as a standard platform for distributed computation has changed this state of a airs. Incentives have become an increasingly important consideration in network protocol design (see, for example, [FNY89, FS97, HA88, KLO95, KS89, S88, S90, S95]) More recently, the work of Nisan and Ronen [NR01] has inspired the design of algorithmic mechanisms for, e.g. scheduling, load balancing, lowest cost paths, and combinatorial auctions that satisfy both the traditional economic de nitions of incentive compatibility and the traditional TCS de ....

Friedman, E., and Shenker, S. (1997). \Learning and Implementation in the Internet," preprint. Available at http://www.icir.org/shenker/decent.ps

....mechanism leads to games in which the Nash equilibrium is unique, robust to coalitional deviations and the only rationalizable strategy pro le. In addition, this Nash equilibrium is the unique outcome of adaptive learning [7] and reasonable learning in asynchronous low information environments [5]. I would like to thank Herve Moulin and Scott Shenker for helpful comments and discussions and Caltech for its hospitality during this work. This material is based upon work supported by the National Science Foundation under Grant No. ANI 9730162. email: friedman econ.rutgers.edu, www: ....

....is analyzed in [11] There are a wide variety of queuing problems which are mathematically equivalent to this problem, such as the ordering of web server or database requests and the allocation of dial in modems. 3 They also arise in a variety of service facilities problems [6, 15, 3] 2 See [5] for an extensive bibliography of such models. 3 These are discussed in detail in [2] 5 In all of these queuing models, the marginal cost is typically increasing in all variables, due to stochastic congestion e ects, and thus our results are applicable. A di erent class of examples arise in ....

[Article contains additional citation context not shown here]

E. J. Friedman and S. Shenker. Learning and implementation in the Internet. mimeo, 1998.

....a cost function C we have imposed no restrictions on the relationship between x(q; C) and x(q 0 ; C) even if q and q 0 are very close. For example, one 4 Shenker [19] and Moulin and Shenker [17] have shown that games induced by serial cost are dominance solvable, while Friedman and Shenker [9] have shown that such games are also solvable in overwhelmed actions, and thus are robustly learnable for a wide class of learning algorithms and information structures. The generalized serial cost method [8] has similar axiomatic and strategic characterizations. 8 might want to require that ....

E. J. Friedman and S. Shenker. Learning and implementation in the Internet. mimeo, 1998.

....many more. We provide simple and easily computable conditions under which these games are learnable by several models of learning, such as adaptive and sophisticated learning (Milgrom and Roberts [34] calibrated learning (Foster and Vohra [8] and reasonable learning (Friedman and Shenker [15]) Thus, using these methods one can evaluate the stability of such games on the Internet or in other settings in which players must learn over time. Running head: Learnability in Internet Games I would like to thank Michelle Goman, Adam Landsberg, Herv e Moulin, and Scott Shenker for helpful ....

.... to the set of Nash equilibria, many dynamic models have been shown to converge the serially undominated set under restrictive informational assumptions, for synchronous play (see, e.g. 34] and more generally to the serially unoverwhelmed set 1 in noisy and asynchronous environments [15]. Thus, certain games are inherently learnable, while others are not. In this paper we consider a large class games which arise on the Internet for which it is easy to determine their degree of learnability. We restrict ourselves to the class of nonatomic games, which are model games with many ....

[Article contains additional citation context not shown here]

E. J. Friedman and S. Shenker. Learning and implementation in the internet. mimeo, 1997.

....no matter how other agents behave whether selfish, spiteful, or stupid truthful revelation is the optimal (dominant) strategy for each user. Strategyproofness is much more exacting than Nash implementation, and so Nash implementation can achieve a much wider variety of outcomes. However, [7] argues that the strategyproofness may be the only viable approach in the Internet, because one cannot ensure that simultaneous selfish play will reach the traditionally defined notions of equilibrium. There is a long history of applying game theoretic techniques to networking problems. Some use ....

E. Friedman and S. Shenker, "Learning and Implementation in the Internet," preprint, 1997. Available at http://www.aciri.org/shenker/decent.ps

....where agents share network bandwidth, as described in [44] is perhaps the most studied example of a repeated game in a network context. As the Internet continues to grow, and more resources are shared by remote users, we expect the network context to become increasingly common. As discussed in [17], the network context di ers from the traditional gametheoretic context in four important ways. I. First, agents have very limited a priori information. In general, agents are not aware of the underlying characteristics of the shared resource. In other words, they do not know the payo structure ....

....traditional concepts of serially unoverwhelmed strategies (O 1 ) and serially Stackelberg undominated strategies (S 1 ) which are discussed below. Our ndings suggest that the asymptotic play of games in network contexts can be quite di erent from that of standard contexts, 5 As discussed in [17], standard control theoretic results imply that the frequency at which strategies are updated should not be greater than the inverse of the round trip communication delay to the shared resource; otherwise, instability may result. 6 Although our present focus is solely on automated agents, ....

[Article contains additional citation context not shown here]

E. Friedman and S. Shenker. Learning and implementation on the Internet. Mimeo, 1997.

No context found.

E. J. Friedman and S. Shenker. Learning and implementation in the Internet. mimeo, 1998.

No context found.

FRIEDMAN, E., AND SHENKER, S. Learning and implementation on the internet. In Manuscript. New Brunswick: Rutgers University, Department of Economics (1997).

No context found.

E. J. Friedman and S. Shenker. Learning and implementation on the Internet. Working paper, 1997.

No context found.

Friedman, E., and S. Shenker (1997). "Learning and Implementation in the Internet," preprint. Available at http://www.aciri.org/shenker/decent.ps

No context found.

E. Friedman and S. Shenker. Learning and implementation on the Internet. Working Paper, 1997.

No context found.

Eric Friedman and Scott Shenker. Learning and implementation on the internet. To appear.

No context found.

E. J. Friedman and S. Schenker. Learning and implementation on the internet. 1998.

No context found.

E. J. Friedman and S. Shenker. Learning and implementation on the internet. Preprint, 1998. (Cited on page 233.)

No context found.

E. J. Friedman and S. Schenker. Learning and implementation on the internet. 1998.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC