An Enhanced Classifier System for Autonomous Robot Navigation in Dynamic Environments

In many cases, a real robot application requires the navigation in dynamic environments. The navigation problem involves two main tasks: to avoid obstacles and to reach a goal. Generally, this problem could be faced considering reactions and sequences of actions. For solving the navigation problem a complete controller, including actions and reactions, is needed. Machine learning techniques has been applied to learn these controllers. Classifier Systems (CS) have proven their ability of continuos learning in these domains. However, CS have some problems in reactive systems. In this paper, a modified CS is proposed to overcome these problems. Two special mechanisms are included in the developed CS to allow the learning of both reactions and sequences of actions. The learning process has been divided in two main tasks: first, the discrimination between a predefined set of rules and second, the discovery of new rules to obtain a successful operation in dynamic environments. Different experiments have been carried out using a mini-robot Khepera to find a generalised solution. The results show the ability of the system to continuous learning and adaptation to new situations.