New Adaptive Multi-Memetic Global Optimization Algorithm for Loosely Coupled Systems

Abstract

This study introduces a new parallel adaptive multi-memetic population-based algorithm for solving global optimization problems efficiently on loosely coupled computing systems. The existent approaches to the synthesis of adaptive population-based algorithms were considered along with the parallelization techniques; distinct features of the loosely coupled computing systems were identified; those features have to be considered carefully when designing algorithms for this class of systems. The proposed algorithm is based on the two level adaptation strategies. The upper level is a static one and allows one to adjust numeric values of the basic algorithm's free parameters before the optimization process using the information about an objective function obtained by means of the landscape analysis. The lower level is a dynamic one and was implemented by means of hybridization of the basic algorithm and several local search methods (memes). The work also presents a new static load balancing method for mapping the proposed algorithm onto parallel computing nodes. The proposed load balancing method takes into consideration both the optimization algorithm's distinct features and the computing system's architecture. This results in higher efficiency of the optimization algorithm comparing to the traditional load balancing methods. A wide performance investigation of the proposed optimization algorithm and its software optimization was carried out in this work with a use of high-dimensional benchmark functions of various classes.