ZCS and TCS Learning Classifier System Controllers on Real Robots

To date there has only been one implementation of Holland’s Learning Classifier System (LCS) on real robots. In this paper the use of Wilson’s ZCS system is described for an obstacle avoidance task. Although the task is simple it does present some advances and change of emphasis over the previous LCS robotic implementation. The controller model is “event” based. Instead of the robot being assigned fixed length actions, continuous actions are taken. These actions are taken until an “event” occurs. An event can be thought of as a change of state. This division of the world into states is usually part of the problem description, and to do this automatically is currently one of the challenges facing machine learning. The paper then introduces TCS, a form of ZCS that attempts to address this issue. LCS have the ability to generalise over the state action-space. In TCS this generalisation ability can also be used to determine the extent of this space. TCS also implements components from SMDP reinforcement learning theory to weight the influence of time on the reward functions of the LCS. A simple light-seeking task on the robot platform using TCS is presented which demonstrates desirable adaptive characteristics for the use of LCS on real robots.