Batch-mode decision tree learning applied to intelligent reactive robot control

摘要

This paper presents an efficient approach based on a symbolic method, namely the decision tree learning, to navigate intelligently a robot in a cluttered, unknown and dynamically changing environment. The two major behaviours, namely reactivity and goal-seeking behaviours, are learned from positively reinforced robot motions from a starting point with no rules. The learning emphasis is on the automatic generation of knowledge without human intervention, with the robot being trained successively to generate knowledge increments in the form of vector entities. A decision tree network is grown on the batch of knowledge fragments to generate coherent decision rules incorporating the behaviours to navigate the robot. We also demonstrate the feasibility of behavioural decomposition into behaviour-biased decision trees.