On the mechanisms of imitation in multi-agent systems

Abstract

Imitation is a social learning method in which an individual observes and mimics another's actions. To implement imitation on robots, a number of questions should be answered, including what information should be copied during imitation, how to choose the behaviors to be copied and how to translate the observed behaviors. In this research, we aim to answer the first two questions in an experiment scenario with simulated agents. First, based on the content of information that is copied during imitation, we compare two different imitation methods, namely, imitation of actions only and imitation of actions and perceptions. It is shown that if the observed behaviors are highly context specific, imitating perceptions along with actions is beneficial compared to imitating actions only. Second, to answer the question of which behaviors to copy, we compared different selection strategies. It is shown that the agents can choose which behaviors to copy by checking the utility of observed behaviors by a trial and error mechanism.