Neural Network architectures design by Cellular Automata evolution

Abstract

The design of the architecture is a crucial step in the successful application of a neural network. However, the architecture design is basically, in most cases, a human experts job. The design depends heavily on both, the expert experience and on a tedious trial-and-error process. Therefore, the development of automatic methods to determine the architecture of feedforward neural networks is a field of interest in the neural network community. These methods are generally based on search techniques, as genetic algorithms, simulated annealing or evolutionary strategies. Most of the designed methods are based on direct representation of the parameters of the network. This representation does not allow scalability, so to represent large architectures very large structures are required. In this work, an indirect constructive encoding scheme is proposed to find optimal architectures of feed-forward neural networks. This scheme is based on cellular automata representations in order to increase the scalability of the method.