Mutual Information As a Task-Independent Utility Function for Evolutionary Robotics

The design of the control system for a swarm of robots is not a trivial enterprise. Above all, it is difficult to define which are the individual rules that produce a desired swarm behaviour without an a priori knowledge of the system features. For this reason, evolutionary or learning processes have been widely used to automatically synthesise group behaviours (see, for instance, Mataric 1997; Quinn et al. 2003; Baldassarre et al. 2007). In this paper, we investigate the use of information-theoretic concepts such as entropy and mutual information as task-independent utility functions for mobile robots, which adapt on the basis of an evolutionary or learning process. We believe that the use of implicit and general purpose utility functions—fitness functions or reward/error measures—can allow evolution or learning to explore the search space more freely, without being constrained by an explicit description of the desired solution. In this way, it is possible to discover behavioural and cognitive skills that play useful functionalities, and that might be hard to identify beforehand by the experimenter without an a priori knowledge of the system under study. Such task-independent utility functions can be conceived as universal intrinsic drives toward the development of useful behaviours in adaptive embodied agents.