A neuro-adaptive congestion control scheme for round trip regulation

A novel neuro-adaptive congestion controller is presented, capable of regulating the per packet round trip time (RTT) around a piecewise constant desired RTT, thus achieving almost piecewise constant delay. The controller is implemented at the source and is proven robust against modeling imperfections, exogenous disturbances (UDP traffic) and delays (propagation, queueing). The notion of communication channels is introduced for throughput improvement. The analysis is nonlinear and the tools used are approximation-based control and linear-in-the-weights neural networks. The proposed controller is guaranteed to be saturated. Moreover, modifications are also provided to achieve rate reduction whenever congestion is detected. Simulation studies illustrate the performance of the proposed control scheme and compare it with other well-established congestion control mechanisms.