Epistasis in Multi-Objective Evolutionary Recurrent Neuro-Controllers

This paper presents an information-theoretic analysis of the epistatic effects present in evolving recurrent neural networks. That is, how do the gene-gene interactions change as the evolutionary process progresses from an initially random state to the final generation and does this reveal anything about the problem difficulty. Also, to what extent does the environment influence epistasis. Our investigation concentrates on multi-objective evolution, where the major task to be performed is broken into sub-tasks which are then used as our objectives. Our results show that the behavior of epistasis during the evolutionary process is strongly dependant on the environment. The experimental results are presented for the path following robot application using continuous-time and spiking neuro-controllers.