Recursive training for multi-resolution fuzzy min-max neural network classifier

Abstract

A new training algorithm for the Fuzzy Min-Max Neural Network (FMMNN) is proposed. The FMMNN model is a powerful tool for pattern classification problems, and is perfect for hardware implementation. But the original model has several unwilling properties. Among them a serious one is how to decide the crucial training parameters. This paper proposes a recursive training algorithm to alleviate the difficulty, and improves the training procedure highly automatic. The result model is a multi-resolution combined classifier (MRCC). Experiments are made following some recent evaluation criteria known in literature, and show that compared with the original model, the MRCC has better classification performance, better adaptive learning ability and consume less computation resource.