Population-based coordinate descent algorithm with majority voting
Davood Zaman Farsa, Azam Asilian Bidgoli, Ehsan Rokhsatyazdi, Shahryar Rahnamayan · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2021
Many real-world optimization problems belong to the class of expensive problems and the costly process of computing fitness value or gradient of the objective function may cause the failure of various optimization algorithms to solve them quickly. Because of the low computation and memory requirements of Coordinate Descent (CD) search methods they are suitable algorithms to optimize these problems. Despite the efficiency of CD methods, searching a large-scale search space just by using one candidate solution decreases the exploration capability of the algorithm. In this paper, a novel population-based version of the CD algorithm called Population-Based Coordinate Descent (PBCD) is proposed which is an efficacious method for tackling such problems using the collective intelligence and collaboration of the population. It takes advantage of three phases of locating the region of interest, folding the search space, and communication among the population members with majority voting to find more promising regions in the search space. As it shrinks the search space swiftly, it needs a low computational budget for finding the optimal value per coordinate and ultimately in overall. To investigate its performance, we benchmarked it on CEC-2017 test suite consisting of 29 low-scale problems with dimensions of 30, 50, and 100.