On the Robustness of Random Walks for Fitness Landscape Analysis

Abstract

Fitness landscape analysis is used to characterize search landscapes of optimization problems. The objective or fitness function of the optimization problem is extended into a search landscape. Random walk algorithms are the most common methods used to sample these search landscapes and various analyses are performed on the obtained samples in order to quantify the fitness landscape characteristics. Random walks are used since they are computationally inexpensive relative to optimization algorithms. However, since random walks are samples of the search landscape, it is important to have the measures that result from the random walks be reliable. This paper investigates the robustness of fitness landscape analysis measures and the effects of higher dimensions on the robustness. This paper suggests that certain random walk algorithms produce robust fitness landscape measures with shorter random walks than other random walk algorithms. Furthermore, this paper suggests that the fitness landscape measures that result from a particular random walk algorithm become robust at similar lengths of random walks across all the optimization functions analyzed.