Coordinators Autonomy Module Technical Report

Coordination of a team of agents assigned to a time critical task has proven to be a challenge for the AI community. If team members include humans, this task becomes even more complex due to human cognitive limitations. To remedy this, humans are usually provided with personal assistants, called proxies, which handle all team level coordination. Since real tasks are actually performed by humans, the interaction between a human and its proxy, called adjustable autonomy, has been an intensive area of research. Three adjustable autonomy challenges have emerged to be particularly important: (i) learning optimal transfer of control strategy between human and its proxy, (ii) learning the human preference model and (iii) learning the human responsiveness model. In this paper we describe AutonMod, a novel adjustable autonomy module which successfully addresses the abovementioned challenges. AutonMod uses Markov Decision Process to learn the different transfer of control sequences; to learn the human preference and responsiveness model it leverages the concept of Neural Networks guided by rules extracted from a given Knowledge Base. We test AutonMod in the scope of DARPA Coordinators project. Our experiments show that AutonMod learns human preference model and increases its confidence in autonomous decision making. We also demonstrate that AutonMod transfer of control strategies outperform näıve adjustable autonomy strategies.