Training support vector data descriptors using converging linear particle swarm optimization

It is known that Support Vector Domain Description (SVDD) has been introduced to detect novel data or outliers. The key problem of training a SVDD is equivalent to solving constrained quadratic programming problem. The Linear Particle Swarm Optimization (LPSO) is developed to optimize linear constrained functions, which is intuitive and simple to implement. However, premature convergence would be followed with the LPSO. The LPSO is extended to the Converging Liner PSO (CLPSO), which is always guaranteed to find at least a local optimum. A new method using CLPSO to train SVDD was proposed. Experimental results demonstrated that the proposed method was feasible and effective for SVDD training, and the performance of it was better than that of traditional method.