Parallel exploitation in estimated basins of attraction: A new derivative-free optimization algorithm

Abstract

Direct search (DS) and evolutionary algorithms (EAs) are two of the most representative branches of derivative-free optimization methods. However, traditional DS becomes deficient in multimodal problems, while EAs suffer from long computational time due to the blind search caused by randomness in evolutionary operators. This paper proposes a new derivative-free optimization algorithm that addresses both the above issues, avoiding prematurity while maintaining fast convergence speed. The new algorithm first estimates basins of attractions in the search space by analyzing samples of the objective function. An adaptive exploitation method with the ability to predict promising search directions is then applied to search the estimated basins in parallel. The new algorithm is evaluated on both unimodal and multimodal benchmark functions. Experimental results show that the algorithm is a promising global optimizer with fast convergence speed. Copyright 2011 ACM.