A comparative simulation study of TCP/AQM systems for evaluating the potential of neuron-based AQM schemes

In recent years, several neuron-based active queue management (AQM) schemes have been proposed. Such schemes exhibit important attributes including fast convergence with high accuracy to a desired queue length. This paper presents extensive comparative simulation results for four neural AQM schemes, namely, Neuron PID, AN-AQM, FAPIDNN, NRL, versus three traditional AQM schemes (ARED, REM and PI) together with a modified PI scheme named IAPI over a wide range of conditions and scenarios. For all schemes, we test their performance in various environments. Through extensive numerical comparisons, we demonstrate that the neuron-based schemes generally achieve faster convergence to queue length target, with smaller queue length jitter. We further demonstrate one order of magnitude reduction in the standard deviation of the end-to-end packet delay by the four neuron-based schemes over the other schemes, when the queueing delay is the dominant delay component. These advantages of neuronal schemes deserve recognition despite the fact that no proof of stability is available for such schemes.