Incremental Aquisition of Complex Visual Behaviour using Genetic Programming and Robot Shaping
In recent years, learning and evolutionary methods have been proposed as methods for automatically designing robot controllers without the need for detailed human design effort. Unfortunately, the reality has been that these methods have only been successfully applied to relatively simple problems involving low-bandwidth sensors and actuators, and simple (often purely reactive) behaviours. Purely automated design methods seem unable to `scale up' to design controllers for the realistically complex tasks we wish to tackle. A promising compromise solution is the idea that the learning/evolutionary system can be left to do most of the work, but with a human providing some sort of high-level assistance to make the problem tractable. Designing robot controllers in this way is often called `robot shaping'. In this thesis I explore a number of dierent forms of shaping, focusing in particular on `black box' techniques which I suggest are more likely to scale up to complex problems than other shaping methods. I also propose a novel extension of Genetic Programming, for use with these shaping methods. Experiments are described in which controllers were evolved, both with and without shaping, for a range of complex tasks including getting a mobile camera to track a moving light in two dimensions, and the harder problem of visually tracking arbitrary moving objects. These controllers are evolved rst in simulation, and then the best ones, evolved using shaping, are transferred successfully to a real robot. I conclude that if used carefully, shaping can reduce learning time and improve nal controller performance. However, choosing an appropriate form of shaping still requires the designer to be very much aware of the underlying details of the evolutionary system. As a result, huma...