Training Reinforcement Neurocontrollers Using the Polytope Algorithm.pdfVIP

TRAINING REINFORCEMENT NEUROCONTROLLERS USING THE POLYTOPE ALGORITHM Aristidis Likas and Isaac Lagaris Department of Computer Science 8 University of Ioannina 9 9 P.O. Box. 1186 - GR 45110 Ioannina, Greece 1 c e D Correspondence: A. Likas 3 Department of Computer Science ] E University of Ioannina N. P.O. Box. 1186 - GR 45110 Ioannina, Greece s c tel: +30-651-97310 [ fax: +30-651-48131 1 v e-mail: arly@cs.uoi.gr 2 0 0 2 1 Abstract 8 9 A new training algorithm is presented for delayed reinforcement learn- / s ing problems that does not assume the existence of a critic model and c : employs the polytope optimization algorithm to adjust the weights of the v i action network so that a simple direct measure of the training performance X is maximized. Experimental results from the application of the method to r the pole balancing problem indicate improved training performance com- a pared with critic-based and genetic reinforcement approaches. Keywords: reinforcement learning, neurocontrol, optimization, polytope al- gorithm, pole balancing, genetic reinforcement. 1 TRAINING REINFORCEMENT NEUROCONTROLLERS USING THE POLYTOPE ALGORITHM Abstract A new training algorithm is presented for delayed reinforcement learn- ing problems that does not assu