Structural Optimization Methods Of Hierarchical Learning Model For Population Diversity

In general,the hierarchical learning model is a kind of method characterized by tree hierarchy structure and automatic evolution,which shows excellent performance in many problems because of its hierarchical and flexible structure.However,the hierarchical learning model is often challenged by the reduced diversity of population during the model evolution.Moreover,the population diversity is a critical element of natural selection in the biological theory and is used to describe the structural or behavioral diversity of populations in genetic programming.Although there is no widely accepted definition of the term diversity,the implicit assumption of it is genotypic diversity,or structural diversity.Therefore,it is reasonable and highly desired to increase the diversity of population through the structure optimization,then improve the individual population searching ability of the hierarchical learning model since the performance of the model is influenced by its structure.This paper proposed a series of researches on the structural optimization of the hierarchical learning model.Pioneers have developed many hierarchical learning models,and two classical models are selected as the main research objects,namely flexible neural tree(FNT)and genetic programming algorithm(GP)in this paper,because FNT and GP show superior performance in classification,and prediction,et al.The main distributions of this paper are summarized as the following aspects:(1)A similarity assessment model(SEFNT)based on FNT is proposed to maintain the diversity of the population and deal with the non-equilibrium problem.There are four contributions in this model.Firstly,according to the specific instructions of FNT,the difference calculation method between different nodes is introduced.Secondly,a nonlinear model is established based on the weight of nodes and the position of nodes in the tree structure,and the similarity distance of the tree structure is calculated.Thirdly,considering the effect of height difference between trees,the amplification factor is adopted.Finally,the non-equilibrium fitness function is used to control its evolution process,so that it can effectively deal with the non-equilibrium problem.Experimental results show that the proposed method not only significantly improves the classification performance of FNT,but also has certain competitiveness compared with other methods.Statistical hypothesis testing results also show that this method has significant advantages.(2)A similarity assessment model(SDCGP)based on GP is proposed to delete similar individuals and guide crossover operations,so as to achieve the goal of increasing population diversity and solve the symbolic regression problem.In SDCGP,firstly,node distance coefficient matrix is introduced to reflect the difference between nodes according to GP instruction set.Secondly,according to different node pairs,the similarity distance calculation function is defined.Then,because GP has some special function instructions(such as +,*),which leads to the symmetry of operation,the way to determine the corresponding node route is proposed.Finally,the adaptive sampling distance is proposed to further ensure the effectiveness of the crossover.The experimental results show that the model can effectively increase the population diversity and the ability to explore the problem space.(3)In this paper,SEFNT model and SDCGP model are applied to the actual problems of video traffic identification and network traffic prediction respectively.A complete framework system for collecting,preprocessing,feature extraction,data collection,identification and prediction of Internet traffic is established.Experimental results show that the proposed algorithm can achieve ideal results in network traffic identification and prediction.In summary,this paper proposes an effective scheme to deal with the structural optimization problems from the two classical hierarchical learning models.experiments verify the effectiveness of the proposed models,and the research work in this paper has theoretical significance and practical value for the structural optimization of the hierarchical learning model and the identification and prediction of Internet traffic.