A Hybrid Q-learning Algorithm to Score a Moving Ball for Humanoid Robots

Abstract

In this paper, we investigate an effective and robust algorithm as a hybrid Q-learning algorithm to instant kick of a moving ball to score the goal. In this regard, as an application, we lied proposed algorithm's characteristics on the humanoid soccer robots. In this paper, we propose a hybridization of Heuristically Generating Possible Actions algorithm (HGPA) and Q-Learning algorithm. In each decision-making process first, the HGPA generates suitable possible actions, and then an adopted Q-Learning algorithm trains and selects the best action. The main objectives of HGPA are focused on the speeding up the learning process of the Q-learning algorithm, and it improves the quality of decision-making process. The simulation results show superior success rates of proposed algorithm in comparison to commonly used Q-learning algorithm in similar problems.