Learning predictive representations

Abstract We demonstrate by a schematic model of an unexperienced animal exploring an environment that it is possible to evolve structures for perception, representation and action simultaneously from a single criterion, namely the error in predicting future sensory inputs. In order to organize successful representations of the environment actions are chosen which are expected to maximize the increase of knowledge. Initially trivial behaviors are generated that allow to learn to recognize places, whereas subsequently virtually random movements indicate that an invariant representation of the environment has emerged.