Comparison with State-of-the-Art: Traps and Pitfalls

摘要

When a new metaheuristic is proposed, its results are compared with the results of the state-of-the-art methods. The results of that comparison are the outcome of algorithms’ implementations, but the origin, names, and versions of the implementations are usually not revealed. Instead, only papers that introduced state-of-the-art are cited. That approach is generally wrong because algorithms are usually described in articles in a way that explains the idea that is hidden behind them but omits the technical details. Therefore, developers have to fill in these details, and they might do so in different ways. The paper shows that even implementations made by one author who is the creator of an algorithm give results which differ considerably from one another. Therefore, for the comparison purpose, the best possible implementation should be identified and used. To illustrate how details that are hidden in the code of implementations influence the quality of the results, sources of quality differences are tracked down for selected implementations. It was found that sources of the differences are hidden in auxiliary code and also stem from implementing a different version of the algorithm which undergoes development. These findings imply best practice recommendations for researchers, implementation developers, and authors of the algorithms.