Q-Learning based on dynamical structure neural network for robot navigation in unknown environment

Abstract

An automation learning and navigation strategy based on dynamical structure neural network and reinforcement learning was proposed in this paper. The neural network can adjust its structure according to the complexity of the working environment. New nodes or even new hidden-layers can be inserted or deleted during the training process. In such a way, the mapping relations between environment states and responding action were established, and the dimension explosion problem was solved at the same time. Simulation and Pioneer3-DX mobile robot navigation experiments were done to test the proposed algorithm. Results show that the robot can learn the correct action and finish the navigation task without people's guidance, and the performance was better than artificial potential field method. © 2009 Springer Berlin Heidelberg.