The Role of Nearly Neutral Mutations in the Evolution of Dynamical Neural Networks

The evolution of continuous time recurrent neural networks is increasingly being employed to evolve nervous systems for autonomous agents. Nonetheless, the picture of populations engaged in hill-climbing rugged fitness landscapes poses a problem of becoming trapped on a local hilltop. Developments in evolutionary theory and molecular biology have pointed to the importance of selective neutrality. The neutral theory claims that the great majority of evolutionary changes are caused not by Darwinian selection but by random drift of selectively neutral or nearly neutral mutants. However, with a few exceptions neutrality has generally been ignored in artificial evolution. This paper addresses the distribution of fitness effects of new mutations when evolving dynamical systems and provides evidence of an improved evolutionary search process when incorporating nearly-neutral drift. This is one of the most fundamental problems in artificial evolution, because it lies at the heart of maintaining a constant-innovative property.