Analysis of Neurocontrollers Designed by Simulated Evolution

Abstract

Randomized adaptive greedy search, using evolutionary algorithms, offers a powerful and versatile approach to the automated design of neural network architectures for a variety of tasks in artificial intelligence and robotics. In this paper we present results from the evolutionary design of a neuro-controller for a robotic bulldozer. This robot is given the task of clearing an arena littered with boxes by pushing boxes to the sides. Through a careful analysis of the evolved networks we show how evolution exploits the design constraints and properties of the environment to produce network structures of high fitness. We conclude with a brief summary of related ongoing research examining the intricate interplay between environment and evolutionary processes in determining the structure and function of the resulting neural architectures. 1. Introduction Artificial neural networks offer an attractive paradigm for the design of behavior and control systems in robots and autonomous agents f...