Competitive Multiagent Systems: Bridging the Gap between Theory and Practice

竞争多智能体系统是两个或多个自利智能体相互作用的环境。正是这种自利因素使多智能体系统具有“竞争性”：不能依赖于多智能体的欲望重合。博弈论是研究这种环境的主要方法。不幸的是，理论和实践之间仍然存在着巨大的差距。例如，博弈论模型通常非常小，建立在代理人无所不知(没有错误信念)和效用最大化(计算无界)的假设之上。即使这样的代理人确实存在，现有的理论也不会告诉他如何对更有限的竞争对手进行推理。资源分配的方法通常假定待售商品不能转售，而且是独一无二的(即，它们只在正在设计的一次性市场中可用)，代理人没有交易费用或预算，市场可以不受任何限制地重新设计，代理人知道他们对所分配商品的估值。在令人感兴趣的实际环境中，这些假设中的一些甚至全部可能无法成立。

Competitive multiagent systems are environments in which two or more self-interested agents interact. It is this ingredient of self-interest that makes a multiagent system ``competitive": agents' desires cannot be relied upon to coincide. Game theory is the dominant approach used to study such environments. Unfortunately, a significant gap still exists between theory and practice. For example, game-theoretic models are typically very small and are built on the assumptions that agents are omniscient (hold no false beliefs) and utility-maximizing (computationally-unbounded). Even if such an agent did exist, existing theories would not tell him how to reason about more limited competitors. Approaches to resource allocation usually assume that goods for sale cannot be resold and are unique (i.e., they are available only in the one-shot market being designed), that agents face no transaction costs or budgets, that markets can be redesigned without any constraints, and that agents know their valuations for the goods being allocated. In practical settings of interest, some or even all of these assumptions can fail to hold.