Multiple Classifier Systems: Software Engineered, Automatically Modular Leading to a Taxonomic Overview

摘要: One vigorous branch of research aimed at improving the performance pattern recognition systems explores possibilities for exploiting differences between a set variously configured classifiers. This is field Multiple Classifier Systems (MCS), and it based on premise that ought to be possible organise exploit strengths weaknesses individual classifiers such MCS superior any its components. Important concerns are efficiency multiple classifier construction, effectiveness final MCS. What property or properties being exploited by various decision strategies, how desired realised within classifiers? Analogous ideas strands have arisen both software engineering neural computing. paper surveys these other two fields from an perspective with goal revealing useful results should direct application current work in In particular, survey opens up new as well provides formal bases central underlying ideas, independence diversity. The exploration diversity extended consideration MCSs which component specialised classification identifiable subset complete problem. Results given empirical study automatic specialisation strategy demonstrates predictive use several measures. Finally, taxonomy presented unifying framework many varieties MCSs.