An artificial three‐layer neural network is constructed to study possible mechanisms in auditory frequency discrimination. The 110 input units, resembling the peripheral neurons, have two sets of filter shapes: idealized eighth nerve isointensity functions and Patterson‐style filters [Moore and Glasberg, Hear. Res.28, 209–255 (1987)]. Twenty‐four intermediate units, similar to critical bands, process information from the input units. The output unit makes a frequency discrimination decision based upon an excitation difference in the intermediate units. Single‐band and multiband decision‐making rules are tested. When using the idealized eighth nerves as the peripheral filter shape, neither the single‐band nor a linear‐combination, multiband model satisfactorily predicts experimental data. However, by adjusting the weights from the intermediate units to the decision‐making unit by the Delta Rule [Rumelhart and McClelland,Parallel Distributed Processing 1(MIT, Cambridge, MA, 1988), pp. 444–459], it is found that a weighted‐band model can best predict the observed data. The weighted‐band model assigns different weights to each band according to its excitation difference: The larger the difference is, the greater the weight. When using Patterson‐style filters as the input units, none of the above three models provides a satisfactory prediction. The results suggest that (a) the peripheral filter shape may play a more important role than the central decision rule in frequency discrimination, and (b) the weighted‐band model may be a more appropriate central decision rule.