Research On The Application Of Semi-supervised Learning Techniques In Expensive Evolutionary Optimization Algorithms

Many optimization problems in the real world are not able to be given in an explicit mathematical formulation and are computationally expensive for simulation.Using cheap surrogate models instead of expensive function evaluations has gained wide attention to assist evolutionary algorithms in searching for the optimal solution.However,the number of training samples is an important factor in ensuring the performance of the surrogate model.Currently,in most optimization algorithms,surrogate models are trained on only a small amount of labelled data to approximate the original functions,making it difficult to train a high-quality model to guide the algorithm to search for the optimal solution.Thus,this thesis proposes to use semi-supervised learning techniques to obtain a set of unlabelled data,which will be merged with labelled data to build an ensemble surrogate to improve the diversity of the model and its performance as well.The main works are given in the following:(1)A semi-supervised learning ensemble surrogate-assisted evolutionary optimization algorithm(SSL-ES)is proposed.At first,two RBF surrogate models are built using two different kernel functions on the labelled data.Simultaneously,a model is built based on the dataset blended with the labelled and unlabelled data.Two models will be chosen for approximating the original problem after comparing the approximated values of solutions by the models trained on the blended data and on the labelled data only.In addition,to further improve the quality of the model,a solution with the best approximated value or with the maximum approximation uncertainty will be adaptively selected by comparing the best solution at the current generation,and the best solution found so far.The CEC2005 benchmark test problem set and a real problem spread-spectrum radar Polly phase code design are used to test the performance of the proposed method.The experimental results show that our proposed algorithm has a significantly better performance compared to some state-of-the-art algorithms.(2)A tri-training-driven ensemble surrogate-assisted evolutionary algorithm(TT-ES)is proposed.Based on the initial exploration of semi-supervised learning methods to build models,an ensemble model is proposed by using the tri-training technique in machine learning to approximate the original optimization problem.Three hybrid datasets are formed by fusing unlabelled data three times to train three RBF surrogate models,where every two models are integrated to approximate the original problem in order to improve the diversity of approximation.A number of experiments are conducted on the CEC2005 benchmark problem set and the spread spectrum radar Polly phase code design problem to compare the performance of the proposed method to some state-of-the-art surrogate-assisted algorithms.The experimental results show that the proposed algorithm has a good performance.