BACKPROPAGATION IN SUPERVISED AND REINFORCEMENT LEARNING FOR MOBILE ROBOT CONTROL

Rudolf Jakša - Peter Sinčák - Pavol Majerník

   The paper deals with the application of a backpropagation algorithm in both, supervised and reinforcement learning approaches in task of the mobile robot navigation. The experimental environment used in both cases is the same. The control is based on a sensor information concerning the position of a vehicle in the environment and a radar information about the obstacles and provide a steering signal and signals for acceleration/deceleration of the vehicle. The control task is to reach the desired position from any point of the environment in the reinforcement learning case and to reach the desired position using supervisor's instructions in the supervised learning case. The neurocontroller consists from two neural networks with the backpropagation learning algorithm accomplishing the reinforcement learning approach or from one simple neural network with the backpropagation algorithm in the supervised learning experiments. The core of this paper was presented at CIMCA'99 Computational Intelligence for Modelling, Control and Automation, (Vienna 1999) conference.

Keywords: backpropagation algorithm, learning approaches, mobile robot navigation, sensor, neurocontroller, neural networks

[full-paper]