Multiple Expert System Design by Combined Feature Selection and Probability Level Fusion

We propose a novel design philosophy for expert fusion by taking the view that the design of individual experts and fusion cannot be solved in isolation. Each expert is constructed as part of the global design of a final multiple expert system. The design process involves jointly adding new experts to the multiple expert architecture and adding new features to each of the experts in the architecture. We evaluate the performance of different fusion strategies ranging from linear untrainable strategies like Sum and Modified Product to linear and nonlinear trainable strategies like logistic regression, single layer perceptron and radial basis function classifier. We investigate two distinct design strategies which we refer to as parallel and serial. In both cases we show that the proposed integrated design approach leads to improved performance.