The application of a unified Bayesian stopping criterion in competing parallel algorithms for global optimization

AbstractThe unconstrained global programming problem is addressed using a multistart, multialgorithminfrastructure, in which different algorithms compete in parallel for a contribution towards a single global stopping criterion, denoted the unified Bayesian global stopping criterion.The use of different algorithms is motivated by the observation that no single (global) optimization algorithm consistently outperforms all other algorithms when applied to large sets of problems from different classes.The Bayesian stopping criterion is based on the single assumption that the probability of each algorithm converging to the global optimum is at least as large as the probability of convergence to any other local minimum. This assumption is often valid in the case of practical problems of physical origin (e.g., determining physical configurations corresponding to minimum potential energy). Results for parallel clusters of up to 128 machines are presented