演化算法中一种基于单分类的预选择策略

In evolutionary algorithms, a preselection operator is a part of reproduction procedure that aims to choose some promising candidate offspring solutions for a further environmental selection. The preselection operator can help to improve the algorithm performance significantly if correctly utilized, in which case the unpromising candidate solutions can be discarded before the real function evaluations and thus the computational resources can be saved. Also, using the preselection operator in evolutionary algorithms will help the algorithms to generate potentially good solutions in one generation without adding more function evaluations. Most existing preselection operators are based on fitness evaluations, surrogate models, and classification models. Since a preselection operation can be regarded as a classification procedure, where selected solutions can be treated as 'positive' ones and the discarded solutions are 'negative' ones. In terms of this situation, using the classification model to assist preselection is a natural choice for evolutionary algorithms. Previous research uses binary and/or multi-class classification models to guide preselection, in which 'positive' and 'negative' training samples or more classes of samples should be prepared to build the classification model. However, after several generations, for some of the evolutionary algorithms, almost all of the solutions in the current population are relatively 'positive' ones, thus the gap between 'positive' and 'negative' solutions is not easy to define. Furthermore, for these kinds of evolutionary algorithms, preparing 'negative' training points has three disadvantages: (1) to reduce the accuracy of the classification model on prediction, (2) to improve the cost on model building, and (3) to increase the complexity of algorithms. For this reason, it is not trivial to prepare 'positive' and 'negative' training samples. To deal with this problem and avoid the above disadvantages, we consider employing the one-class classification (OCC) model, which only needs one class of 'positive' training samples in the classification model build to guide preselection. Thus, the model-building procedure can become simple by only defining the 'positive' training points. Based on this idea, this paper proposes a one-class classification based preselection (OCPS) scheme that uses the OCC model for the preselection. The proposed OCPS scheme mainly has three components: at first, solutions in the current population are all labeled as 'positive' ones with label+1. Then, the labeled solutions are used to build an OCC model. Thirdly, for each parent solution, a set of candidate offspring solutions are generated. The built model is employed to label the newly generated offspring solutions and only the one with 'positive' label will be selected as the real offspring solution of its parent for the following environmental selection. If there are more than one candidate solutions are labeled as 'positive' or there are not any solutions labeled as 'positive', the real offspring solution will be randomly selected from these newly generated candidates. The proposed OCPS scheme is applied to three state-of-the-art evolutionary algorithms on a test suite. The experimental results show the potential of OCPS on improving the performance of some existing evolutionary algorithms and saving the number of function evaluations. Also, the OCPS performs better than the other models in most situations.