Abstract

Multiagent systems are increasingly being used to solve a wide variety of problems in a range of applications such as distributed sensing, information retrieval, workflow and business process management, air traffic control and spacecraft control, amongst others. Each of these systems has to be designed at two levels: the micro-architecture level, which involves the design of the individual agents and the macro-architecture level which involves the design of the agents' organizational structure. In this research, we are primarily concerned with the agents' macro-architecture.

At the macro-architecture level, the multiagent designer is concerned with issues such as the number of agents needed to solve the problem, the assignment of tasks to the agents and the coordination mechanisms being used. The design of the agents' macro-architecture is complicated by the fact that there is no best way to organize and all ways of organizing are not equally effective. Instead the optimal organizational structure depends on the problem at hand and the environmental conditions under which the problem needs to be solved. In some cases, the environmental conditions may not be known a priori, at design time, in which case the multi-agent designer does not know how to develop an optimal organizational structure. In other cases, the environmental conditions may change requiring a re-design of the agents' macro-architecture. These are only a few of the many hurdles confronting the macro-architecture designer.

In our research, we simplify the macro-architectural design by passing on some of the macro-architectural design responsibilities to the agents themselves. That is, instead of manually designing the macro-architecture of a multiagent system at design time, we allow the agents to come up with their own organizational structure at run time. This approach is known as Organizational Self Design (OSD) and it allows the agents to adapt their organizational structure to changing environmental conditions and differences in the problems being solved.

Most of the current work on OSD has focused on task-oriented domains. In our research, we extend OSD to apply to worth-oriented domains, the hardest class of problems. Our research focuses on developing algorithms and mechanisms that allow (a) the generation of agents as an artifact of the system; and (b) the generation of different organizational structures that make different quality/cost tradeoffs based on the organizational design constraints specified and the performance criteria being optimized. Such tradeoffs are not possible in task-oriented and state-oriented domains.