ABSTRACT
In most research using reinforcement-learning techniques, task and re ward are tightly coupled: The task for the learning agent is to maximize its reward. There are many cases however, especially in multirobot sys tems, where the task and reward must be treated separately. Autonomous agents embedded in their environment are not always able to accurately ac cess their performance, and system-wide performance may depend on other agents over which the learner has no direct control or communication. The
taxonomies of task and reward presented here provide a framework for in vestigating the impact of differences in the performance metric and reward on multirobot system performance.