Group behavior and group learning

摘要

We describe an approach to principled synthesis and analysis of group behavior in situated, embodied multiagent systems. We propose basic behaviors as the appropriate level for control and learning. Basic behaviors are generated by simple local rules and serve as building blocks for a large repertoire of higher level behaviors. We describe an architecture for combining basic behaviors into compound, more complex tasks. We also describe a formulation of reinforcement learning that allows for learning such compound behaviors automatically in non Markovian, noisy and uncertain environments with multiple agent. We demonstrate all of our methodologies with experimental data on a collection of physical mobile robots demonstrating group avoidance, aggregation, dispersion, following, wandering, flocking, and foraging.