Research On Selection And Optimization Of Ensemble Learning Based On Multimodal Multiobjective Optimization

Ensemble learning is a method to complete machine learning tasks by constructing different basic learners.In recent years,ensemble learning occupies an important position in machine learning.Accuracy and diversity are two conflicting objectives to be optimized in ensemble learning.Generally speaking,improving accuracy will make diversity worse and vice versa.It is a challenging task to balance both two objectives.The main task of this research is to achieve the best trade-off between them.Especially as the widespread use of swarm intelligence algorithms,some scholars improve the generalization ability of ensemble learning.One of the reason is that evolutionary algorithms can optimize both the combination of strategy weight parameters and selection of base learners.However,there are also some disadvantages.Some algorithms cannot consider the two objectives to build an ensemble learning model,and cannot fully find the equivalent optimal solution.It is necessary to improve the ensemble learning models.To deal with the above issues,this thesis proposes an ensemble learning based on multimodal multiobjective optimization.The main components of the proposed method are: 1)Establish an ensemble learning model selectively;2)Optimize the connection weights inside the base learner to build an ensemble learning model.The algorithm uses a neural network as the base learner,codes the parameters of the base learner and base learner separately,and builds an ensemble learning model as well as considering both accuracy and diversity.Since the function of multimodal multiobjective optimization algorithm is to obtain many equivalent optimal solutions.Then these equivalent optimal solutions are used to build good ensemble learning models.The main contributions are listed as follows.1.A basic learner selection algorithm based on multimodal multiobjective optimization is proposed.The neural network is used as the base learner,and accuracy and diversity are taken as the two objectives.The base learner is coded,and then experiments is performed on classification data sets.The experimental results also show the effectiveness and significance of using multimodal multi-objective optimization to solve ensemble learning problems.Multimodal multiobjective optimization can provide more choices for the construction of ensemble learning models,without reducing classified accuracy of ensemble learning in the overall process.2.A parameter optimization algorithm for base learners based on multimodal multiobjective optimization is proposed.Neural networks are used as the base learner,which encodes the internal connection weights of the base learner and optimizes the weights of the evolutionary algorithm.The performance of the algorithm is verified on classification data sets.Experimental results show the effectiveness of the algorithm and the necessity of using a multimodal multiobjective optimization.Also the built ensemble learning model has a good generalization ability.The instability of the parameters in the base learner due to randomness is solved.The proposed algorithm is compared with other classical algorithms,and experimental results show that the proposed algorithm can achieve better results on most data sets.As pointed out above,the thesis can build a good ensemble learning model by considering the both accuracy and diversity.It has been verified that the research of ensemble learning is of great significance.The classification and summary of the combination of evolutionary algorithms and ensemble learning are given,since the optimization problem of ensemble learning has a multimodal multiobjective attribute,it is necessary to use multimodal multiobjective optimization approaches.