Evolving modular fast-weight networks for control

Abstract

In practice, almost all control systems in use today implement some form of linear control. However, there are many tasks for which conventional control engineering methods are not directly applicable because there is not enough information about how the system should be controlled (i.e. reinforcement learning problems). In this paper, we explore an approach to such problems that evolves fast-weight neural networks. These networks, although capable of implementing arbitrary non-linear mappings, can more easily exploit the piecewise linearity inherent in most systems, in order to produce simpler and more comprehensible controllers. The method is tested on 2D mobile robot version of the pole balancing task where the controller must learn to switch between two operating modes, one using a single pole and the other using a jointed pole version that has not before been solved. © Springer-Verlag Berlin Heidelberg 2005.