This survey is presented as a description of the field of MAS. It is designed to serve both as an introduction for people unfamiliar with the field and as an organizational framework for system designers. This framework is presented as a series of three increasingly complex and powerful scenarios. The simplest systems are those with homogeneous non-communicating agents. The second scenario involves heterogeneous non-communicating agents. Finally, the general MAS scenario involves communicating agents with any degree of heterogeneity. Single-agent systems are presented as the most extreme version of this final, most complex scenario, where control is centralized in one agent and the others act as remote slaves.
Each multiagent scenario introduces new issues and complications. Although MAS is a new field, several techniques and systems already address these issues. After summarizing a wide range of such existing work, useful future directions are presented. Throughout the survey, Machine Learning approaches are emphasized.
Although each domain requires a different approach, from a research perspective the ideal domain embodies as many issues as possible. Robotic soccer is presented here as a useful domain for the study of MAS. Systems with a wide variety of agent heterogeneity and communication abilities can be studied. In addition, collaborative and adversarial issues can be combined in a real-time situation. With the aid of research in such complex domains, the field of MAS should continue to advance and to spread in popularity among designers of real systems.
MAS is an active field with many open issues. Continuing research is presented at dedicated conferences and workshops such as the International Conference on Multi-Agent Systems [95, 2, 1]. MAS work also appears in many of the DAI conferences and workshops [22, 94]. This survey provides a framework within which the reader can situate both existing and future work.