Designing Parallel Meta-Heuristic Methods

Meta-heuristic methods represent very powerful tools for dealing with hard combinatorial optimization problems. However, real life instances usually cannot be treated efficiently in "reasonable" computing times. Moreover, a major issue in metaheuristic design and calibration is to make them robust, i.e., to provide high performance solutions for a variety of problem settings. Parallel meta-heuristics aim to address both issues. The objective of this chapter is to present a state-of-the-art survey of the main parallel meta-heuristic ideas and strategies, and to discuss general design principles applicable to all meta-heuristic classes. To achieve this goal, we explain various paradigms related to parallel meta-heuristic development, where communications, synchronization and control aspects are the most relevant. We also discuss implementation issues, namely the influence of the target architecture on parallel execution of meta-heuristics, pointing out the characteristics of shared and distributed memory multiprocessor systems. All these topics are illustrated by examples from recent literature. These examples are related to the parallelization of various meta-heuristic methods, but we focus here on Variable Neighborhood Search and Bee Colony Optimization.