Distributed Resource Management Strategies based on Game Theoretic Approaches

◈연  사 : 박 형 곤 박사 (the Swiss Federal Institute of Technology)
◈ 일  시 : 2009년 10월 20일 (화) 오전 11:00 ~ 12:00
◈ 장  소 : LG동 강당
◈ 초청자 : 정 홍 교수 (T.2223)

Abatract
Multimedia applications such as Internet TV, peer-to-peer (P2P) multimedia streaming/broadcasting, and media stream analysis are increasingly deployed over resource constrained systems or network infrastruc-tures such as the Internet, P2P/wireless networks, etc.
To enable the proliferation of these applications, we propose distributed and dynamic multi-user resource management frameworks. This enables users that repeatedly interact with each other in a dynamically varying environment to strategically maximize their own utilities based on their heterogeneous processing abilities. As an illustrative example, we consider P2P networks, where multiple peers are interested in sharing content. We model the resource reciprocation among the peers as resource reciprocation games and show how the peers can determine optimal strategies for resource reciprocation using a Markov Deci-sion Process (MDP) framework. The optimal strategies determined based on MDP enable the peers to make foresighted decisions about resource reciprocation, such that they can explicitly consider both their immediate as well as future expected utilities. We consider heterogeneous peers that have different and limited ability (bounded rationality) to characterize their resource reciprocation with other peers, and analytically study how the bounded rationality can impact the interactions among the peers and the resulting resource reciprocation.
As an alternative illustration, we consider the problem of optimizing stream mining applications, constructed as tree topologies of classifiers, deployed on a set of resource constrained and distributed processing nodes. We design distributed solutions, by defining tree configuration games, where individual classifiers configure themselves to maximize an appropriate local utility. We analytically investigate the convergence of the proposed approach and the corresponding performance. Then, we evaluate the performance of our solutions on an application for sports scene classification, by actually implementing them on the IBM System S processing core middleware.