AbstractThe ART (Adaptive Resonance Theory) neural network was proposed by Grossberg and colleagues to classify by certain criterion (vigilance parameter) a data set distributed across feature space. With ART1, the selection of existing categories and generation of new categories is performed appropriately using two procedures called competition and resonance. Past learning can coexist with new learning without contradictions (that is, the so‐called plasticity–stability dilemma is avoided). However, ART classification results are influenced by the data presentation order, so that different numbers of categories and different classification results may be obtained for the same data set, which is basically unacceptable. As a solution to this problem, this study introduces two procedures, namely, asymptotic setting of the vigilance parameter and probabilistic restoration of prototypes. Asymptotic setting of the vigilance parameter is a procedure to reduce the number of generated categories to a minimum by varying the vigilance parameter gradually from its minimum to a preset target value; in the process, the classification results develop from rough to fine. Probabilistic restoration of prototypes handles changes in the order of data presentation by probabilistic variation of existing prototypes (representative values). By adding these two procedures, an ART modification may be designed with a reduced number of generated categories, while not being dependent on data presentation order. In this study, the two additional procedures were introduced into ART1, variety of ART for dealing with binary patterns, and the effectiveness of the proposed algorithm was verified against several data sets. Experiments proved that the proposed method provides order‐independent data classification with a minimum (or nearly minimum) number of categories. © 2001 Scripta Technica, Syst Comp Jpn, 32(11): 61–69, 2001