On Learning and Adaptation in Multiagent Systems: A Scientific Computing Perspective
Anupam Joshi, Tzvetan Drashansky, John R. Rice, Sanjiva Weerawarana, Elias N. Houstis · 1995
Systems with interacting agents are now being proposed to solve many problems grouped together under the "distributed problem solving" umbrella. For such systems to work properly, it is necessary that agents learn from their environment and adapt their behaviour accordingly. We investigate such systems in the context of scientific computing. The physical world consists of interacting system, and its overall behaviour emerges from the interacting local behaviours of its constituents.In this paper we present a system which uses a combination of neuro--fuzzy learning and static adaptation to coordinate the activity of multiple agents. An epistemic utility based formulation is used to automatically generate the exemplars for learning, making the process unsupervised. We illustrate how these techniques can be used to convert a standalone, single agent system into a collaborative, multiagent one, and present some results from a preliminary implementation. We also present the design...