Hybrid関数における適応DEの振る舞いについて

摘要

Differential Evolution (DE) is an Evolutionary Algorithm (EA) that was primarily designed for real parameter optimization problems [2]. Despite its relative simplicity, DE has been shown to be competitive with more complex optimization algorithms, and has been applied to many practical problems [3]. As with other EAs, the search performance of DE algorithms depends on control parameter settings [3–5]. A standard DE has three main control parameters, which are the population size N, scaling factor F, and crossover rate CR. However, it is well-known that the optimal settings of these parameters are problem-dependent and parameter tuning in real-world problem is often infeasible for various reasons. Since this is a significant problem in practice, adaptive mechanisms for adjusting the DE control parameters on-line during the search process have been studied by many researchers [5–7]. In this paper, we show that state-of-the-art adaptive DE algorithms such as SHADE [7], JADE [6], and jDE [5] perform extremely poorly on a class of hybrid functions [10] composed of 2 components. An in-depth analysis of the failure of DEs on hybrid functions, and show extensive evidence that adaptive DEs are easily deceived by hybrid functions – they tend to quickly adapt to solve the component of the hybrid function that is easier to make progress on, and are unable to solve the other component. Thus, this study identifies a class of adaptive-DE-hard problems that confound adaptive DEs and poses a challenging direction for future work.