NN's and GA's: evolving co-operative behaviour in adaptive learning agents

Abstract

Without a comprehensive training set, it is difficult to train neural networks (NN) to solve a complex learning task. Usually, the more complex the problem or task the NNs have to learn, the less likely it is that there is a realistic training set that could be used for (supervised) training. One way to overcome this limitation is to implement an evolutionary approach to train NNs. We report the results of a novel implementation of an evolutionary computational technique, based on a modified genetic algorithm (GA), to evolve feedforward NN topologies and weight distributions. The learning task was for two fairly simple but autonomous agents (controlled by NNs) to learn to co-operate in order to accomplish a task. Given the complexity of the task, an evolutionary approach to a search for an appropriate NN topology and weight distribution seems to be a promising and viable approach, as our results show. The implications of the results are further discussed.<>