Evolving neural network controllers

Abstract

An emerging design paradigm uses evolutionary processes to search for optima in design space. The evolutionary technique has the advantage of being a declarative paradigm; the user specifies the task, and a genetic algorithm searches for an optimum solution. Normal techniques require the definition of the controller, and this is computationally expensive. We use a genetic algorithm to design a neural network-based controller for a hexapod robot. The robot must perform the task of moving from a start position to a goal position, under varying degrees of simulated instrument and sensor noise. The findings show that it is possible to embed a degree of noise tolerance into the solution. This is useful in situations where the environment of the robot may change over time.