Virtual bagging for an evolved agent controller

An evolved agent controller is used to produce multiple votes for its choice of actions by presenting it with rotations and reflections of its actual sensory inputs. Since the agent´s behavior with respect to the orientation of its various sensors must be learned independently, the votes can be treated as independently learned opinions on the optimal choice of actions, i.e. a form of Breiman´s bagging. The mechanism is tested on neural controllers for agents trained by neuroevolution in a game-like simulator, and is found to improve their task performance as well as ensuring perfect symmetry of behavior with respect to orientations of the environment.