Statistical Analysis of Multilayer Perceptrons Performances.pdfVIP

Statistical Analysis of Multilayer Perceptrons Performances Remus BRAD, Ioan MIHU and Macarie BREAZU “Lucian Blaga” University of Sibiu Computer Science Department Bulevardul Victoriei 10, 2400 Sibiu, Romania rbrad/mihuz/mac@cs.sibiu.ro Abstract In section 2, we present the framework of our experiments. Section 3 shows the result of our statistical investigations The paper is based on a series of studies on the learning and defines the effective size of the MLP tested. A capabilities of multi-layered perceptrons (MLP). The performance analysis is performed in section 4, by complexity of these nonlinear systems can be varied, acting decomposing error in bias and variance terms. for instance on the number of hidden units, but we will be confronted with a choice dilemma, concerning the optimal 2 Data Sets and MLP Architecture complexity of the system for a given problem. By the mean of statistical methods, we have found that the effective We have employed in all our experiments, training and number of hidden units is smaller than the potential size; validation sets derived from the classification problem some units have a binary activation level or a time known as the Breiman waveforms [3]. This is a constant activation. We also prove that weight initialization classi