Multi-agent based evolutionary artificial neural network for general navigation in unknown environments

Abstract

This paper presents a multi-agent based evolutionary artificial neural network (ANN) for general navigation. While vision is a single input channel to the ANN, only the information about the availability of places in the current visual field is considered so that navigation is executed without restriction to any specific environment or object. Through constant interaction with the environment, multiple agents co-decide and compete with each other for the move decisions. These agents are subject to evolution via evolutionary strategies which are believed to be capable of solving the epistasis problem. Evolutionary learning continues during the lifetime of the ANN and is triggered whenever a correct move decision is absent. These basic ideas are illustrated by a virtual creature exploring unknown environments as far as possible with obstacle avoidance. Experimental results have shown that the virtual creature has incrementally improved its navigation ability and explored the environments efficiently Meanwhile, some natural and robust behaviors have emerged, which are not programmed in advance.