| As the fast progress of Internet technology,a massive amount of data has been produced in various industries.Support vector machine(SVM),one of the most commonly used data classification techniques,has been widely used in many fields because of its unique advantages in the cases of nonlinearity,small samples and "dimensional disaster".However,the parameter choice of this method has a great influence on the performance of the model,so finding the optimal parameters of SVM has become a research topic of interest.Particle swarm optimization algorithm(PSO)is a simple and easy algorithm with fewer parameters and faster convergence,which is currently used by many scholars for SVM parameter optimization,but the algorithm is liable to fall into the problem of local optimum when performing the search.In response to the above deficiencies,three SVM parameter optimization approaches based on modified particle swarm algorithm are proposed in this paper.The main work is as follows.(1)A particle swarm optimization algorithm(CWPSO)with a dynamically varying inertia weight based on a nonlinear inertia weight is studied using Kent chaotic mapping,and then the algorithm is applied to search for the optimal parameters of the SVM.In CWPSO,the use of a dynamically varying inertia weight based on a chaotic mapping facilitates the algorithm to find the optimal solution in the early stage and convergence in the later stage.In addition,when the algorithm falls into local optimum,this inertia weight helps the PSO to jump out of that state.In view of the good convergence accuracy,stability and robustness of CWPSO when tested on the benchmark function,the algorithm is utilized to carry out parameter optimization of the SVM,and the CWPSO＿SVM classification model is built.Finally,experiment is conducted on four UCI data sets,and the numerical result indicates that CWPSO＿SVM has good classification performance.(2)A particle swarm optimization algorithm(CBIWPSO)based on beta distribution and inverse incomplete gamma function is proposed and used to optimize the parameters of SVM.CBIWPSO improves the optimal search capability of PSO in two ways,firstly,by generating a more uniform initial population through a chaotic mapping,thus improving the quality of the initial solution and the likelihood of searching for the optimal solution.Secondly,a dynamic adjustment of inertia weight based on beta distribution and inverse incomplete gamma function is introduced into the velocity update formula,which is conducive to balancing global exploration and local exploitation,and also reduces the possibility of PSO falling into local optimum in the late stage.Since CBIWPSO algorithm has good global exploration and local development ability when optimizing the test function,the algorithm is utilized to search the optimal parameter combination of SVM,and the CBIWPSO＿SVM model is constructed.The result of the model on the data sets shows that the proposed method can improve the classification accuracy.(3)A particle swarm optimization algorithm(DROPSO)based on dimensional random opposition learning is proposed and the SVM parameters are optimized using the DROPSO algorithm.During initialization,the current solution and its random opposition learning solution are searched in parallel to find a better initial solution,so as to lay the foundation for the following global optimization.Secondly,the dimensional strategy is integrated with random opposition learning to perturb the global optimal solution and reduce the probability of the algorithm falling into the local optimum.The experimental result indicates that DROPSO algorithm has been greatly improved in convergence accuracy and stability.Finally,the parameters of the SVM are optimized using the algorithm,and the DROPSO＿SVM classifier is built.The classifier can achieve good prediction accuracy by testing on the UCI data sets. |
