Learning to Behave Socially

摘要

Our previous work introduced a methodology for synthesizing and analyzing basic behaviors which served as a substrate for generating a large repertoire of higher--level group interactions (Matari'c 1992, Matari'c 1993). In this paper we describe how, given the substrate, agents can learn to behave socially, i.e. to maximize average individual by maximizing collective benefit. While this is a well--defined problem for rational agents, it is difficult to learn in situated domains. We describe three sources of reinforcement and show their necessity for learning non--greedy social rules. The learning strategy is demonstrated on a group of physical mobile robots learning to yield and share information in a foraging task. 1 Introduction Our previous work focused on analyzing and synthesizing complex group behaviors from simple social interactions between individuals (Matari'c 1992, Matari'c 1993). We introduced a methodology which involved designing a collection of basic behaviors which se...