Social learning for collaboration through ASiCo based neuroevolution

Abstract

This paper discusses an algorithm that provides a way to obtain ensembles of collaborating artificial neural networks (ANNs) online. That is, its purpose is to find solutions to problems based on the interaction of sets of, in principle, heterogeneous ANNs whose joint behaviour results in an emergent solution. This approach is intrinsically able to handle lifelong adaptation within the society in order to comply with changing situations or demands in dynamic environments. It is called Asynchronous Situated Coevolution (ASiCo) and was designed for the lifelong coevolution of artificial neural network societies. ASiCo deals with the evolutionary part of neuroevolution and it can support any type of neural network structure or even neural network construction mechanism. Consequently, it can be extended with some of the techniques found in single ANN neuroevolutionary mechanisms when considering the simultaneous evolution of network weights and network topology. The operation and characteristics of this strategy are illustrated through some experiments carried out using a well known benchmark collaboration task.