Partitioned multiagent systems in information oriented domains
Claudia V. Goldman, Jeffrey S. Rosenschein · 1999
Abstract Multiagent systems often operate in environments where access to, and manipulation of, information is paramount. We formally characterize a subset of these domains as Information Oriented Domains, and consider a novel organizational approach to handling multiagent problem solving in these environments. The initial step is to partition the information domain, using an evolutionary algorithm and a similarity criterion. This partitioning step results in a group of agents, each with a particular area of "expertise", assigned to handle the available information. Because of the nature of the evolutionary algorithm, the number of final agents and their assignments arise naturally from the information content of the environment, and are not determined a priori. This organization of agents, spanning the information space, can then be exploited in problem solving and information retrieval tasks. The partitioning of the information space leads to greater efficiency at run time. In general, these partitioned systems are appropriate for information retrieval systems, or (more specifically) for libraries of software tools or partial plans. Moreover, these organizations can also be used as the basis for domain-specific problem solvers, where the partitioning leads to efficiency in searching the solution space. As an illustration of this last advantage of dynamically partitioned systems, we describe an organizationbased multiagent planner that we have designed and implemented. The agent expertise covers a set of HTML documents, and makes use of semantic tags added to these documents. A user can query the system, retrieving both the answer desired as well as additional related