Evolutionary robot behavior via natural selection based on neural networks

The traditional fitness function based methodology of artificial evolution is argued to be inadequate for the construction of entities with behaviors novel to their designers. Evolutionary emergence via natural selection (without an explicit fitness function) is the way forward. This paper primarily considers the question of what to evolve, and focuses on principles of developmental modularity in neural networks. To develop and test the ideas, an artificial world containing autonomous organisms has been created and is described. Experimental results show that the developmental system is well suited to long-term incremental evolution. Novel emergent strategies are identified both from an observer´s perspective and in terms of their neural mechanisms.