Research and Design of Intelligent Many-Agent Systems

摘要

This report documents efforts under ONR grant no. N00014-94-1-0676. This is anAASERT award attached to parent grant NRL no. N00014-93-1-0022. The purpose of the grant is to support research on how group dynamics can emerge from collections of agents that would enable them to make decisions that individuals could not or accomplish tasks that individuals could not. Funding from the grant supported four graduate students directly; i.e., with stipends and tuition, and a number of undergraduate students indirectly, through materials and supplies purchases to support their independent study efforts in distributed intelligence and cooperative robotics. Results of these studies indicate that among distributed/cooperative learning methods, the most promising and appropriate for distributed mobile agent applications is a combination of learning and behavioral methods. In particular, the recommended method combines the data structures and execution cycle of the learning classifier system with reinforcement computed similarly to Q-learning and with some stochastic selection and genetics-based rule-paring methods. These systems, in conjunction with message-based communications between agents, is shown to be widely applicable and convergent in ideal scenarios. The methods have the disadvantages of being slow, and they do not perform well in sequential learning tasks without significant modifications.