Design and Analysis of a novel weightless artificial neural based Multi-Classifier
Pierre Lorrentz, W.G.J. Howells, Klaus Dieter McDonald-Maier · Kent Academic Repository (University of Kent) · 2007
Recent years have witnessed intense research in the general area of Multi-Classifier sy stems (MCS), but this has rarely incorporated any utilisa tion of weightless neural systems(WNS) as the combiner of an MCS ensemble. This paper explores the application of weightless networks within the multi-classifier environment by introducing an intelligent multi- classifier system using a WNS called the Enhanced Probabilistic Convergent Neural Networks (EPCN). The paper explores the use of EPCN by illustrating its ma jor features, such as the specification of disjoint or overlapping input subset to the MCS, and the inherently parallel nature of the design. Within the proposed system, the number of base classifiers per MCS could be specified manually or automatically. . The proposed MCS is problem-domain independent and, our investigation is performed on handwritten characters. The proposed MCS is adaptive, its combiner is capable of extracti ng absolute or weighted classification decision(output ) from base classifier. Diversity is increased in the base classifier by injecting randomness into the system parameters. Two types of EPCN classifiers are proposed, fix-PCN and rand-PCN. These PCNs are independent and orthogonal. One uses a fixed method of forming connectivity while the other uses random method of forming connectivity. In order to verify the performance of the recogni tion system, tests were performed, off-line, on benchmark datasets of unconstrained handwritten numerals. Experimental results suggest that MCS outperforms single EPCN in classification of handwritten characte rs.