Authors: Shu Ying-Li; Ke Wen-De
Addresses: School of Electrical Engineering and Information Technology, Changchun Institute of Technology, Changchun, China ' Department of Computer Science, Guangdong University of Petrochemical Technology, Guangdong, China
Abstract: A novel switching group-size genetic algorithm (GA) based on Hierarchical Reinforcement Learning (HRL) is proposed to improve the global optimisation performance by accelerating the convergence speed of genetic algorithm and improving the computational efficiency. In the early stage of the evolution, the group can be expanded to increase diversity, while the group should be downsized to protect the more adaptive individuals in the latter stage. Chromosome crossover operation in the HRL algorithm is regarded as behaviour. Choosing the optimal method according to the specific evolution of the chromosome reflects the optimisation selection. At the same time, abstract and hierarchical system is used to decompose the problem to multilevel sub-task space. The convergence speed is improved by learning strategically in each sub-task space and multiplexing the sub-strategy between layers. The experiment has proved the validity of the algorithm.
Keywords: genetic algorithms; hierarchical reinforcement learning; HRL; convergence speed; global optimisation.
International Journal of Wireless and Mobile Computing, 2016 Vol.11 No.4, pp.277 - 282
Received: 24 Jun 2016
Accepted: 28 Sep 2016
Published online: 14 Feb 2017 *