In previous work , an incremental radial basis function network trained by a dynamic decay adjustment algorithm (RBFNDDA) was integrated with histogram to reduce redundant hidden neurons (or simply neurons). In order to remove unnecessary neurons, a weight-based indicator was utilized . This hybrid model (RBFNDDA-HIST1) can reduce unnecessary neurons and maintain classification accuracy satisfactorily. However, another aspect of noises, i.e., overlapping among neurons of different classes in RBFNDDA-HIST1 and RBFNDDA, is not tackled fully for solutions. To close this research gap, another version of RBFNDDA-HIST (i.e., RBFNDDA-HISTR) is developed whereby the radius of a neuron (that overlaps with neurons of other classes) is checked before removing it from the network. Several public data sets that have a high level of overlapping records according to an overlapping indicator are used to evaluate the performance of RBFNDDA-HISTR in terms of number of neurons and classification accuracy. A performance comparison among RBFNDDA, RBFNDDA-HISTR and RBFNDDA-HIST1 are made. The results show that the proposed RBFNDDA-HISTR can reduce the number of neurons from RBFNDDA-HIST1 without deteriorating classification accuracy.