One of the recently emerged paradigms in machine learning is multiple classifier fusion. A large number of methods for constructing multiple classifier systems (MCS) have been suggested in the literature. The majority of these draw on a parallel architecture, involving a fusion of multiple classifiers via some form of linear or nonlinear combination rule. Intuitively, one can look at parallel fusion as an attempt to improve the performance by combining several independent estimates of a class aposteriori probability and thereby reducing the variance of the combined estimate. For a given probability margin between two competing hypotheses, this reduced variance then results in a lower probability of incurring an additional classification error over and above the Bayes' error. Much less attention has been paid to multiple classifier system schemes that aim to enhance the performance by manipulating the margin between competing hypotheses. An increased margin can normally be achieved by class grouping. This approach often leads to serial multiple classifier system architectures. Depending on whether the grouping structure is fixed or created dynamically, the resulting multiple classifier is either a decision tree or a chain like multistage system. In this paper the theory underpinning this MCS approach will be overviewed and its implications discussed. It will be shown that the theory leads to diverse class grouping/margin manipulation strategies. Their relative advantages will be discussed. The effectiveness of some of these strategies will be illustrated on a practical problem of object recognition.