A comparative study of decision combination strategies for a novel multiple-expert classifier

The performance of a novel multiple expert decision combination strategy has been compared with other multiple expert decision combination methods reported in the literature. The concept of decision combination has been generalised in two different categories and it has been demonstrated how these different categories perform with respect to each other under optimised conditions. The paper presents the performance of this particular network, which is a Type-II network and compares it with other Type-I decision combination strategies previously reported in literature. These methods include aggregation method, choice selection and ranking method. In all the cases, the chosen database was the NIST database, which is recognised to be the standard database for handwritten characters. It has been found that this particular Type-II configuration is able to outperform all these Type-I combination strategies. The performance enhancement on a subset of the NIST database having a thousand character samples for each class has been found to be around 1.2% with respect to the best recognition performance obtained from either of the Type-I decision combination strategies investigated.