Behavior Analysis Of Evolutionary Computing Via Unsupervised Feature Learning

As a kind of heuristic optimization method,Evolutional Computing(EC)has achieved great success in dealing with concrete complex optimization problems in the last decades.However,theoretical analysis of EC is very difficult due to its complex random behav-ior.No effective methods have been proposed till now to learn or analyze the behavior of evolutionary algorithms under different circumstances.For better understanding the behavior of EC,we utilized some unsupervised feature learning algorithms in this article to analyze the behavior of one generation behavior during the search process.Firstly,we give the definition of the behavior data of EC for behavior data constructing.Mainly we used three algorithms to complete the feature extraction and behavior analysis task:Self-organized map(SOM)based behavior data pre-processing,Slow Feature Analysis(SFA)based feature extraction and Deep Belief Networks(DBN)based feature extrac-tion.The main contribution or work of this article can be listed as below:1)Research on the SOM based evolutionary computing behavior data pre-processing methods.The t-SNE based pre-training method was discussed for training SOM net-work.Thus the training process can be divided into three parts:pre-training,rough-training and fine-tune,which guarantee the SOM network converge to the best state.Then we use the trained SOM neural network to map the original high-dimensional be-havior data to a 2D plain,for the purpose of achieving the normalized representation of high dimensional data set,which were used for latter behavior analysis.2)Research on the SFA based EC behavior data feature extraction algorithms.Firstly,we adjusted the time structure in the original form of SFA to fit for our unsuper-vised pattern recognition problem.Also,we explored the dimension choosing problem of the slow features.Then we designed a 2-order nonlinear expansion based SFA fea-ture extraction framework according to the characteristic of the evolutional computing behavior.Finally,several comparison experiments were designed with the purpose of finding out whether there exist differences between the searching behavior of different evolutionary algorithms running on the same fitness landscape and whether there are dif-ferences between the behavior of one algorithm running on different fitness landscapes.The experimental results show that SFA can extract robust features to discriminate dif-ferent evolutionary algorithms.3)Research on the DBN based EC behavior data feature extraction algorithms.We firstly made a deep exploration on RBM(Restricted Boltzmann Machine),which is the basic component of DBN.Then according to the evolutional computing behavior,we designed a DBN structure,who contained 7-layer RBMs.As the above,we evaluated the DBN based feature extraction algorithm through several experiments,and visual-ized the features of different evolutionary algorithms' behavior data on the same test function.Besides,we also made a comparison to SFA and analyzed the behavior of four different evolutionary algorithms.