An ancestor based extension to Differential Evolution (AncDE) for Single-Objective Computationally Expensive Numerical Optimization

Abstract

This paper presents the Ancestral Differential Evolution (AncDE) algorithm, which extends the standard Differential Evolution (DE) algorithm by adding an archive of recently discarded ancestors. AncDE adds the ability to occasionally compute difference vectors between current and archived solutions, using these inter-generational difference vectors in place of traditional difference vectors. Results for AncDE are presented for the CEC2015 Bound Constrained Single-Objective Computationally Expensive Numerical Optimization Problems using AncDE/best/1/bin. Summary results are included for standard DE for comparison purposes and these show that AncDE generally outperforms standard DE. These results suggest that the inter-generational difference vectors can help overcome some local optima, leading to faster convergence towards the global optimum. AncDE involves the very small overhead of storing and updating the ancestral cache. This paper introduces two empirically determined stochastic rates; one for updating the ancestral cache and the other for using an ancestral difference vector in place of the normal difference vector.