Agents Modeling Agents in Information Economies

Our goal is to design and build agents that act intelligently when placed in an agent-based information economy, where agents buy and sell services (e.g. thesaurus, search, task planning services, etc.). The economy we are working in is the University of Michigan Digital Library (UMDL), a large scale multidisciplinary effort to build an infrastructure for the delivery of library services [2]. In contrast with a typical economy, an information economy deals in goods and services that are often derived from unique sources (authors, analysts, etc.), so that many goods and services are not interchangeable. Also, the cost of replicating and transporting goods is usually negligible, and the quality of goods and services is difficult to measure objectively: even two sources with essentially the same information might appeal to different audiences. Thus, each agent has its own assessment of the quality of goods and services delivered. Our emphasis, therefore, is not on developing market mechanisms for traditional economies with interchangeable goods/services, but rather for those where each participant might be unique and inscrutable from the perspective of others. In order to make good decisions, an agent in such an economy must: (A) Determine exactly what services other agents provide and at what price. (B) Avoid agents whose services are not reliable or needed, and form teams with those that provide needed services. (C) Decide how much to charge and whom to target. We believe these can be accomplished if an agent builds models of how others appear to be assessing quality and/or establishing prices, and even how others are modeling others in these ways. Our previous work [l] considered such recursive models, and gave algorithms that trade-off the time costs of using the deeper recursive models versus the costs of taking a possibly inferior action. However, it assumed the agents already had correct deeper models of others. The next problem to solve is how agents can ac-