Extending Minimum Population Search towards large scale global optimization

Abstract

Minimum Population Search is a new metaheuristic specifically designed for optimizing multi-modal problems. Its core idea is to guarantee exploration in all dimensions of the search space with the smallest possible population. A small population increases the chances of convergence and the efficient use of function evaluations — an important consideration when scaling a search technique up towards large scale global optimization. As the cost to converge to any local optimum increases in high dimensional search spaces, metaheuristics must focus more and more on gradient exploitation. To successfully maintain its balance between exploration and exploitation, Minimum Population Search uses thresheld convergence. Thresheld convergence can ensure that a search technique will perform a broad, unbiased exploration at the beginning and also have enough function evaluations allocated for proper convergence at the end. Experimental results show that Minimum Population Search outperforms Differential Evolution and Particle Swarm Optimization on complex multi-modal fitness functions across a broad range of problem sizes.