EVOLVING COMPLEX VISUAL BEHAVIOURS USING GENETIC PROGRAMMING AND SHAPING

Abstract

Introduction 1.1 From Robot Learning. . . For many years now, automatic design approaches based on techniques from Genetic Algorithms, Reinforcement Learning and Neural Networks, have been touted as the way in which the engineers of the future will produce robot control systems. Just specify at a high level what you want the robot to do, and then sit back and wait while the GA/RL/NN works out how to get the robot to do it. While this is undoubtedly a very attractive goal, a quick look at the current state of the art of automated robot design reveals that in most cases we are a long way from being able to automatically design controllers that can out-perform human-designed ones. While robot controllers have been learned from scratch for very simple `insect-like' behaviours, general-purpose tabula rasa learning of complex tasks seems di#cult or impossible. 1.2 . . . To Robot Shaping One particularly promising solution to this problem is to bring the human back into the design cycle: