Siti Sendari, Shingo Mabu, Kotaro Hirasawa
This paper proposes Two-Stage Reinforcement Learning on Credit Branch Genetic Network Programming named GNP-TSRL-CB for mobile robots. The proposed method uses 2 kinds of Q-tables for sub node selection and credit branch selection, which has advantages of (1) determining an alternative function by using sub node selection and (2) skipping useless functions by using credit branch selection. It is clarified from simulation results that the adaptability mechanism of the proposed method can improve the performance compared with the conventional methods when the individuals of GNP-TSRL-CB are implemented in the dynamic environments like the sudden changes occur. © 2013 The Institute of Electrical Engineers of Japan.
State University of Malang, Malang, East Java, 65145, Jalan Semarang No. 5, Indonesia; Graduate School of Information, Production, and Systems, Waseda University, Wakamatsu-ku, Kitakyushu, Fukuoka 808-0135, 2-2, Hibikino, Japan
