Confidence combination methods in multi-expert systems

In the proposed paper, we investigate the combination of the multi-expert system in which each expert outputs a class label as well as a corresponding confidence measure. We create a special confidence measurement which is common for all experts and use it as a basis for the combination. We develop three combination methods. The first method is theoretically optimal but requires very large representative training data and storage memory for look-up table. It is actually impractical. The second method is suboptimal and reduces greatly the required training data and memory space. The last method is a simplified version of the second and needs the least training data and memory space. All three methods demand no mutual independence of the experts, thus should be useful in many applications.