Rating Players in Games with RealValued Outcomes

12th International Conference on Autonomous Agents and Multiagent Systems

(Extended Abstract) ; Matthew Rutherford

PID  :  123933

来源: CEUR

【 摘 要 】

Gametheoretic models typically associate outcomes with real val ued utilities, and rational agents are expected to maximize their expected utility. Currently fielded agent rating systems, which aim to order a population of agents by strength, focus exclusively on games with discrete outcomes, e.g., winloss in twoagent settings or an ordering in the multiagent setting. These rating systems are not wellsuited for domains where the absolute magnitude of util ity rather than just the relative value is important. We introduce the problem of rating agents in games with realvalued outcomes and survey applicable existing techniques for rating agents in this setting. We then propose a novel rating system and an extension for all of these rating systems to games with more than two agents, showing experimentally the advantages of our proposed system. Categories and Subject Descriptors I.2.1 [Artificial Intelligence]: Application and Expert Systems— Games; I.2.11 [Artificial Intelligence]: Distributed Artificial In telligence—Multi Agent Systems General Terms Experimentation, Economics, Measurement