Surrogate Model-Based Differential Evolution

Abstract

In many practical applications, optimization problems have a demanding limitation on the number of function evaluations that are expensive in terms of time, cost and/or other limited resources. Adaptive differential evolution such as JADE is capable of speeding up the evolutionary search by automatically evolving control parameters to appropriate values. However, as a population-based method by nature, adaptive differential evolution still typically requires a great number of function evaluations, which is a challenge for the increasing computational cost of today’s applications. In general, computational cost increases with the size, complexity and fidelity of the problem model and the large number of function evaluations involved in the optimization process may be cost prohibitive or impractical without high performance computing resources. One promising way to significantly relieve this problem is to utilize computationally cheap surrogate models (i.e., approximate models to the original function) in the evolutionary computation.

In this chapter, JADE is extended to address optimization problems with expensive function evaluations by incorporating computationally inexpensive surrogate models and adaptively controlling the level of incorporation according to model accuracy. Radial basis function networks are used to create the surrogate models for the sake of good balance between computational complexity and model accuracy. Experimental results have shown the efficiency of adaptive incorporation of surrogate models in avoiding potential false or premature convergence while significantly reducing the number of original expensive function evaluations.