Learning predictive representations

We demonstrate by a schematic model of an unexperienced animal exploring an environment that it is possible to evolve structures for perception, repre sentation and action simultaneously from a single criterion, namely the err or in predicting future sensory inputs. In order to organize successful rep resentations of the environment actions are chosen which are expected to ma ximize the increase of knowledge. Initially trivial behaviors are generated that allow to learn to recognize places, whereas subsequently virtually ra ndom movements indicate that an invariant representation of the environment has emerged. (C) 2000 Elsevier Science B.V. All rights reserved.