Teaching new teammates

Abstract

Knowledge transfer between expert and novice agents is a challenging problem given that the knowledge representation and learning algorithms used by the novice learner can be fundamentally different from and inaccessible to the expert trainer. We are particularly interested in team tasks, robotic or otherwise, where new teammates need to replace currently indisposed team member(s). We are interested in a general knowledge transfer framework where existing team members or experts can train a new agent to follow its role in team coordination by using exemplars of desirable behavior. Each such exemplar presents a team situation and a preferred action. We envisage an iterative training process where the trainer selects more exemplars in the next iteration based on the errors made by the learner in action choices for test exemplars presented in the current iteration. Such an iterative, exemplar based generic knowledge transfer scheme can be used by agents using arbitrary knowledge representation and learning methods. We evaluate the success of training new teammates in the well-known pursuit problem, where some of the current set of expert predators is being replaced by new ones with no a priori hunting knowledge. Experimental results demonstrate the robustness of our knowledge transfer scheme with a graceful performance degradation.