Engineering Mechanics

Proceedings Vol. 11 (2005)

ENGINEERING MECHANICS 2005

Copyright © 2005 Institute of Mechanics of Solids, Faculty of Mechanical Engineering, Bruno University of Technology, Brno

 

Turek M., Březina T., Pulchart J.
pages 315 - +8p., full text

Modifications of reinforcement learning algorithm, so called continuous action reinforcement learning automaton (CARLA), are presented in this contribution. Automaton learning algorithm is based on rewarding that gradually evolves the set of probability densities. This set is consequently used for action set determination. Modifications consist of improving learning parameters based on learned values. Thereby higher values of probability density near the best action are reached and therefore the variance of chosen actions is lower than original. The influence of modifications is proved by simulation study describing learning and behavior of asynchronous electromotor scalar control. Standard PSD controller is used whose parameter values represent actions of three independent automata. The goal of on line learning process is to minimize the mean square of control error. Here described modifications of algorithm allow the improvement of quality of revolutions control with preserving basic algorithm characteristics.

Text and facts may be copied and used freely, but credit should be given to these Proceedings.

All papers were reviewed by members of the scientific committee.