Improving backtracking search algorithm with variable search strategies for continuous optimization

Abstract

The backtracking search algorithm (BSA), a relatively new evolutionary algorithm (EA), has been shown to be a competitive alternative to other population-based algorithms. To effectively solve a variety of optimization problems, this paper suggests ten mutation strategies and compares the performance of selection mechanisms in employing these strategies. Moreover, following the original BSA design, new parameters of historical mean and best positions are proposed in order to implement several additional mutation strategies. In addition, as recommended in the literature, a one-dimensional crossover scheme is enacted for greedy strategies in order to prevent premature convergence. Furthermore, three settings for search factors of mutation strategies are proposed. As a result, improved BSA versions that employed, respectively, ten and four mutation strategies were found to significantly facilitate the ability of BSA to handle optimization tasks of different characteristics. The experimental results show that the proposed versions outperformed the basic BSA in terms of achieving high convergence speed in the early stage, reaching the convergence precision and plateau with better scores, and performing perfectly on tests of composition functions. In addition, the improved BSA versions outperformed five popular, nature-inspired algorithms in terms of achieving the best convergence precision and performing perfectly on six composition functions.