A New Kind of Fuzzy Neural Network Classifier

陈曦,靳东明,李志坚

DOI: https://doi.org/10.3321/j.issn:0372-2112.2002.06.043

2002-01-01

Abstract:A Multi-Resolution Combined Classifier (MRCC) model and its learning algorithm are proposed.The MRCC is a modification of Simpson's original Fuzzy Min-Max (FMM) Neural Network Classifier.It overcomes some undesired properties of the original model:specifically,training results do not depend heavily on pattern presentation order and hyper-box expansion is not limited by a fixed maximum size.Compared with the original model,a new parameter called the reliability of hyper-box is introduced in MRCC,which presents the correct classification rate of a single hyper-box.A locally uniform distribution hypothesis is also introduced,so that the reliability of hyper-box parameter can be adjusted on-line.In addition,the multi-resolution combination method alleviates the effort to select an optimal parameter for the maximum size of a hyper-box in the original algorithm.Experiments were made following some recent evaluation criteria known in literature,and show that compared with original model,the MRCC model has better classification performance,more adaptive learning ability and creates less hyper-boxes.