The Evolution of Behavior: Some Experiments
Jean-Arcady Meyer, Stewart W. Wilson · 1991
This paper investigates the evolutionary development of (problem solving) behavior. Through evolution, artificial animals learn to survive in a given world. We use layered neural networks (NNs) as the substrate on which evolutionary learning operates. The fault tolerance of neural networks allows for a genotype I phenotype distinction which maintains the variation in the genetic pool. Furthermore, we define building blocks which take into account the functionality of the NNs. The result of the algorithm can be inspected at two levels. First, there is the behavior of the individual animals. A description of their behavior is obtained through the induction of decision trees which describe the functionality of the NN. Second, the behavior of the population as a whole can be described. The distribution of the animals over the world often provides an analogical representation of a problem solution.