Multiagent systems represent a special class of dynamic systems, where the change over time is due to the interaction of multiple agents. Examples of such systems systems abound in the real world and the laboratory. In recent research, popular demonstration applications have included robotic soccer [1,17], the predator-prey pursuit game [5], and networked information systems [26]. Another increasingly salient example is a marketplace, where prices of goods continuously change as agents announce offers and make deals. As the rise of the Internet and electronic commerce continues, dynamic automated markets will be an increasingly important domain for software agents.
In designing agents to participate in an automated market, we face an array of general issues in dynamic multiagent systems. Agents need to consider not only the dynamic changes of the system, but also the reactions of other agents. Whereas agents may have some a priori basis for predicting the behavior of others, in typical situations much further evidence is revealed during the course of interaction. Thus, an agents' learning capacity can play a large role in its success. Depending on the type and quantity of information revealed through the interactions, an agent might learn about the others' joint or individual actions. In either case, this learning must be online because the other agents do not make themselves available for free trials.
In this paper, we adopt regression methods for online derivation of functional relations between other agents' actions and their internal states. The functional form of a regression model is based on an agent's assumptions of other agents' underlying behaviors. In this paper, we show that how an agent can form different models based on its different assumptions about other agents, and how such assumptions relate to the performance of the learning agent.
We define our problem in the framework of dynamic games. After the general discussion, we introduce a dynamic market trading game (interaction through synchronous double auctions), which is an instance of the more general class of dynamic multiagent systems. We propose a series of regression-based learners for this domain, varying in strategy and depth of reasoning about others. In our experiments, learning agents with minimal assumptions about other agents tend to perform better than agents that attribute excess sophistication to their counterparts. However, when such attribution is warranted, agents can do better by learning the more sophisticated models.