Problem-independent Parallel Simulated Annealing using selection and migration

Abstract

Parallelization of the Simulated Annealing algorithm is a difficult task, due to its inherent sequential nature. A popular alternative approach has been to perform a number of independent annealings, often augmented by periodic “coordination”, the generic strategy being to periodically pick the best current solution as the new starting point for all of the continued annealings.

We have conjectured that a more refined strategy may lead to improved performance. We have sought inspiration from Parallel Genetic Algorithms based on the so-called island model. We have implemented three versions of parallel Simulated Annealing; a version without coordination, a version with “traditional” global coordination, and a version with local coordination based on principles from the island model. The experimental results indicate that the proposed local coordination strategy generally outperforms the two other versions.