Multiple Exper t Decision Combination Approach for Character Recognition

Abstract

In an attempt to achieve an agreeable decision, it is crucial to coordinate the opinions provided by different experts, whose expertises usually induce dissimilar conclusions. Classification problems and character recognition problems are one of the well-known instances. These problems need to apply different features to distinguish one from another. However, different features usually provide different suggestions. To avoid this kind of inconsistency, one can apply voting methods or weighting methods to generate a final decision compulsively. Nevertheless, these methods are flawed inherently. Taking voting methods for example, it is not always adequate to resort to the majority since the opinion from an authority sometimes is more credible than those from several ordinary experts. On the other hand, traditional weighting methods that assign constant weight to each expert is insufficient because experts can not only be justified according to the domains they specialize in, but also the targets they treat with. Based on the above observations, we propose a new approach, named MEDC, to support a combined decision among multiple experts. It takes advantage of the philosophy underlying neural networks and refines the traditional weighting method to achieve our goal.