Design of a multiple classifier system

Classification performance can be improved through the use of multiple classifier systems. In this paper, a formulation for the design procedure of a multiple classifier system is considered after compared with a single classifier system. Two phases of the design procedure, i.e., coverage optimization and decision optimization, as well as the techniques for them are reviewed. Although there do exist some methods, the optimization design of a multiple classifier system is still an open issue. A particular application of multiple classifier systems is given which is about type recognition of aircraft in air warfare. In the application, multiple neural network classifiers are constructed through the subspace method, and are combined by the sum rule and stacking to form multiple classifier systems. Experimental results show that the multiple classifier system generated by the sum rule outperforms other classification strategies, while stacking does not work well. These observations are analyzed and probable cause is given.