Real Time Evolution of Behavior and a World Model for a Miniature Robot using Genetic Programming

A very general form of representing and specifying an autonomous agent's behavior is by using a computer language. The task of planning feasible actions could then simply be reduced to an instance of automatic programming. We have evaluated the use of an evolutionary technique for automatic programming called Genetic Programming (GP) to directly control a miniature robot. To our knowledge, this is the rst attempt to control a real robot with a GP based learning method. Two schemes are presented. The objective of the GP-system in our rst approach is to evolve real-time obstacle avoiding behavior from sensorial data. This technique enables real time learning with a real robot using genetic programming, it has, however, the drawback of the learning time being limited by the response dynamics of the environment. To overcome this problems we have devised a second method, learning from past experiences stored in memory. This new system allows speeds up of the algorithm by a factor of more than 2000. The emergence of the obstacle avoiding behavior is also speeded up by a factor of 40 enabling learning of this task in 1.5 minutes. This learning time is several orders of magnitudes faster then comparable experiments with other control architectures. The used algorithm is furthermore very compact and can be tted into the micro-controller of the autonomous mobile miniature robot.