Global optimization with derivative-free, derivative-based and evolutionary algorithms

This paper investigates global optimization methods from the perspective of population-based and restarted point-based heuristics. We examine the performance of a standard evolutionary computation (EC) methodology, a derivative-based sequential quadratic programming (SQP) algorithm and a novel derivative-free stochastic coordinate ascent (SCA) algorithm. All methods are analyzed by random sampling of the feasible search space. A comparison was made to evaluate the three algorithms, in the light of newly updated IEEE CEC2013 benchmarks, on a set of multimodal and composite test cases. Results revealed that while the standard EC algorithm is generally more robust, on the basis of convergence efficiency both the restarted SCA and SQP algorithms have shown remarkable performance on some of these benchmarks. The results further suggest that depending on the nature of the problem landscape and dimensionality the three algorithms, chosen from different optimization frameworks, perform complementary to each other.