ATM call admission control using a neural network trained with a virtual output buffer method

A new adaptive call admission control method that uses a neural network is proposed for ATM communication networks. The neural network is trained using virtual cell-loss data observed from virtual output buffers, which simulate the actual cells being multiplexed into imaginary ATM links of various bandwidths. By interpolating and extrapolating the virtual cell-loss data, the neural network can accurately estimate the cell-loss rate for various bandwidths and traffic loads. This method therefore does not require the observation of actual cell-loss events in a running ATM node. To learn the accurate mean cell-loss rate from widely-distributed observed cell-loss data, the smoothed-log-conversion method is proposed, in which the teacher signal is generated from the weighted sum of the neural-network-estimated cell-loss rate and the data observed at virtual buffers