Incremental acquisition of local networks for the control of autonomous robots

Abstract

This paper proposes an incremental learning approach to control autonomous robots based on local networks. This approach integrates different learning techniques in a conceptually simple architecture. The robot does not learn from scratch, but uses two types of bias: builtin reflexes (domain knowledge) and advice. The robot adds a new unit to the neural network whenever it uses the reflexes or receives an advice. This unit is integrated into a topology preserving map and associates a region around the current situation to either the computed reflex or the advice. The resulting reaction rule is then tuned by means of reinforcement learning and self-organizing rules. Experimental results show that the robot TESEO rapidly learns suitable behavioral strategies.