SVM Hyper-parameters optimization using quantized multi-PSO in dynamic environment

Support vector machine (SVM) is considered as one of the most powerful classifiers. They are parameterized models build upon the support vectors extracted during the training phase. One of the crucial tasks in the modeling of SVM is to select optimal values for its hyper-parameters, because the effectiveness and efficiency of SVM depend upon these parameters. This task of selecting optimal values for the SVM hyper-parameters is often called as the SVM model selection problem. Till now a lot of methods have been proposed to deal with this SVM model selection problem, but most of these methods consider the model selection problem in static environment only, where the knowledge about a problem does not change over time. In this paper we have proposed a framework to deal with SVM model selection problem in dynamic environment. In dynamic environment, knowledge about a problem changes over time due to which static optimum values for yper-parameters may degrade the performance of the classifier. For this there should be one efficient mechanism which can re-evaluate the optimal values of hyper-parameters when the knowledge about a problem changes. Our proposed framework uses multi-swarm-based optimization with exclusion and anti-convergence theory to select the optimal values for the SVM hyper-parameters in dynamic environment. The experiments performed using the proposed framework have shown better results in comparison with other techniques like traditional gird search, first grid search, PSO, chained PSO and dynamic model selection in terms of effectiveness and efficiency.