Multi-Classifier Systems: Review and a roadmap for developers

Multi-Classifier Systems (MCSs) have fast been gaining popularity among researchers for their ability to fuse together multiple classification outputs for better accuracy and classification. At present, there is a lot of literature covering many of the issues and concerns that MCS designers encounters. However, we found out that there isn’t a single paper published thus far which presents an overall picture of the basic principles behind the design of multi-classifier systems. Therefore, this paper presents a current overview of MCSs and provides a road-map for MCS designers. We identify all the key decisions that a designer would have to make over the design of a MCS and list out the most useful options available at each decision making step. We also present a case-study of the MCS theoretical issues considered, and present informal guidelines for the selection of different paradigms, based on the properties and distribution of the data. We also introduce a novel optimization of the standard majority voting combiner which uses a genetic algorithm.