Grammars and cellular automata for evolving neural networks architectures

Abstract

The class of feedforward neural networks trained with back-propagation admits a large variety of specific architectures applicable to approximation pattern tasks. Unfortunately, the architecture design is still a human expert job. In recent years, the interest to develop automatic methods to determine the architecture of the feedforward neural network has increased, most of them based on the evolutionary computation paradigm. From this approach, some perspectives can be considered: at one extreme, every connection and node of architecture can be specified in the chromosome representation using binary bits. This kind of representation scheme is called the direct encoding scheme. In order to reduce the length of the genotype and the search space, and to make the problem more scalable, indirect encoding schemes have been introduced. An indirect scheme under a constructive algorithm, on the other hand, starts with a minimal architecture and new levels, neurons and connections are added, step by step, via some sets of rules. The rules and/or some initial conditions are codified into a chromosome of a genetic algorithm. In this work, two indirect constructive encoding schemes based on grammars and cellular automata, respectively, are proposed to find the optimal architecture of a feedforward neural network.