基于协同过滤的连续黑箱优化问题元启发算法选择

算法选择(AS)问题旨在为给定问题在算法集合中选择最佳算法.随着优化算法的不断提出,算法选择问题是优化领域亟待解决的问题.提出基于聚类的元启发算法五星评价体系,将算法性能指标映射至整数评价以减小评价空间.通过测试24种常见优化算法与4种最新CEC大赛优胜算法在219种、3 000多个标准测试问题上的性能,得到评价矩阵.将评价矩阵作为训练数据,使用协同过滤(CF)算法建立算法评价的预测模型.使用该模型预测算法集内的所有算法在新问题上的评价,结果显示所提出方法预测精度较高,超过90%的预测最佳算法为最终可行算法.敏感性分析显示,该方法在先验信息有限的情况下仍可以保持较高的预测精度.

Metaheuristic algorithm selection system for continuous black-box optimization problems based on collaborative filtering

Selecting the best algorithm out of an algorithm set for a given problem is referred to as the algorithm selection(AS) problems. The importance of AS problems increases with the emerge of many optimization algorithms. Therefore, a five-star ranking system based on clustering is proposed, which maps the algorithm performance criteria to integers and reduces the ranking space. An algorithm set is prepared, including 24 commonly used optimization algorithms and four algorithms that win the CEC competition in 2016 and 2017. By testing the performance of the selected algorithms on 219 benchmark problems, a ranking matrix is obtained. The ranking matrix is used as the data source of the collaborative filtering (CF) algorithm to obtain a prediction model of algorithm ranking. For a new problem instance, the model predicts the ranking of all the algorithms in the algorithm set. The results show that the prediction accuracy is high, and over 90% of predicted best algorithms are capable of solving the problem instance. Sensitivity analysis shows that the proposed method can still maintain high prediction accuracy with limited prior information.