Expeditive Extensions of Evolutionary Bayesian Probabilistic Neural Networks

Abstract. Probabilistic Neural Networks (PNNs) constitute a promising methodology for classification and prediction tasks. Their performance depends heavily on several factors, such as their spread parameters, kernels, and prior probabilities. Recently, Evolutionary Bayesian PNNs were proposed to address this problem by incorporating Bayesian models for estimation of spread parameters, as well as Particle Swarm Optimization (PSO) as a means to select prior probabilities. We further extend this class of models by introducing new features, such as the Epanechnikov kernels as an alternative to the Gaussian ones, and PSO for parameter configuration of the Bayesian model. Experimental results of five extended models on widely used benchmark problems suggest that the proposed approaches are significantly faster than the established ones, while exhibiting competitive classification accuracy.

• Type: Conference
• Year: 2009
• Where: LION
• Authors: Vasileios L. Georgiou, Sonia Malefaki, Konstantinos E. Parsopoulos, Philipos D. Alevizos, Michael N. Vrahatis
• Keywords: Bayesian Model, LION 2009, Optimization, Prior Probabilities, Spread Parameters