A parallel adaptive swarm search framework for solving black-box optimization problems

Abstract

This paper presents a framework to support parallel swarm search algorithms for solving black-box optimization problems. Looking at swarm based optimization, it is important to find a well fitted set of parameters to increase the convergence rate for finding the optimum. This fitting is problem dependent and time-consuming. The presented framework automates this fitting. After finding parameters for the best algorithm, a good mapping of algorithmic properties onto a parallel hardware is crucial for the overall efficiency of a parallel implementation. Swarm based algorithms are population based, the best number of individuals per swarm and, in the parallel case, the best number of swarms in terms of efficiency and/or performance has to be found. Data dependencies result in communication patterns that have to be cheaper in terms of execution times than the computing in between communications. Taking all this into account, the presented framework enables the programmer to implement efficient and adaptive parallel swarm search algorithms. The approach is evaluated through benchmarks and real world problems.