A method of gene selection that decreases the number of fitness value calculations
Satoshi Yamaguchi, Hidekiyo Itakura · Electrical Engineering in Japan · 2003
Abstract This paper proposes a new selection method of a gene from a population for reducing the number of fitness value calculations in genetic algorithms. The proposed method effectively reduces the computation time required to search for a solution to the optimization of a process involving a large number of such calculations as in robot behavior decisions and neural network learnings. In the proposed method, the use of a special buffer for storing a gene and its associated fitness value is introduced. The gene in the buffer is used as a candidate for the solution to the optimization problem. This gene is compared with a gene selected from the population, and one of the values is used in the next generation depending on the results of the comparison. The proportion of suitable genes in the population is increased in the possible shortest time. The convergence speed of our buffer depends on the population topology, which is introduced for choosing a gene from the population to compare with the buffer gene. Three kinds of topology are applied to our algorithm and they are compared and evaluated. The proposed method is applied to a robot control problem to demonstrate the validity of the technique. © 2003 Wiley Periodicals, Inc. Electr Eng Jpn, 143(4): 42–49, 2003; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/eej.10148