CONNECTIONIST Q-LEARNING IN ROBOT CONTROL TASK

摘要

The study deals with the Q-learning algorithm that belongs to reinforcement learning algorithms. The use of neural networks in the Q-learning algorithm and in its modifications - the modified Q-learning and the Q(lambda) algorithm - is considered. Each algorithm is examined in the context of two methodologies enabling one to speed up the process of learning: backward replay and online learning. Efficiency analysis of the algorithms was performed experimentally by means of a software robot simulator. In the course of the experiments, the task of robot control in the continuous environment was solved. The dislocation of obstacles , their configurations and goal dislocation at each stage of learning were changed during the experiments. An analysis of the results obtained is made and the effectiveness of the algorithms is evaluated.