Learning multiple conflicting tasks with artificial evolution

Abstract

The paper explores the issue of learning multiple competing tasks in the domain of artificial evolution. In particular, a robot is trained so as to be able to perform two different tasks at the same time, namely a gradient following and rough terrain avoidance behaviours. It is shown that, if the controller is trained to learn two tasks of different difficulty, then the robot performance is higher if the most difficult task is learnt first, before the combined learning of both tasks. An explanation to this superiority is also discussed, in comparison with previous results.