Robotic systems are characterised by large state and action spaces, which are problematic for learning and adapting techniques. This paper presents a method to alleviate learning dimensionality problems, by classifying similar states and actions into abstract concepts. Unlike previous techniques, concepts are generated based on the robot-environment interaction rather than being arbitrarily imposed. The validity of this method is demonstrated empirically using a real robot platform in a balltargeting task.
|Name||Frontiers in Artificial Intelligence and Applications|
|Conference||Proceedings of the Second Starting AI Researchers Symposium Volume 109 Frontiers in Artificial Intelligence and Applications|
|Period||1/01/04 → …|