Effective modifications to differential evolution optimization algorithm

Abstract

Many inverse problems utilize optimization algorithms to perform minimization of least squares norms in an accurate, reliable and computationally efficient manner. Due to the cost of evaluating real-world objective functions, optimization algorithms must be both fast and robust. Differential Evolution (DE) algorithm is known for its robustness, but not its speed. This paper proposes four simple modifications to DE and compares their performance to Particle Swarm (PS) algorithm using a subset of the Schittkowski & Hock test cases. With these techniques, DE is observed to converge to the global minimum up to three times faster than PS, while maintaining robustness in some cases, and generally performs better than the original forms of DE.