Searching among Search Spaces: Hastening the genetic evolution of feedforward neural networks

Abstract

The paper introduces a genetic paradigm for evolving feedforward neural network, where both the network topology and the weights distribution are coded in the individuals. The resulting binary coded strings are too long for efficient evolution, thus two novel techniques are employed. The first consists in a coding procedure that allows the genetic algorithm to evolve the length of the coding string along with its content. This goes by the name of granularity evolution procedure. The second technique is inspired by linear programming and yields a fast and accurate fine tuning of the solutions.