Towards Satisfying an Application Level Quality Measure via a Neural Network TCP-like Protocol

A new TCP-like protocol is proposed in this paper aiming at regulating the packet round trip to satisfy a user imposed desired application sending time. For this purpose linear in the weights neural networks are employed to construct an on-line model for the estimated round trip depending on source rate (control input) and congestion level which is directly measured from the network via monitoring the state of the ECN bit for each packet acknowledgement. When the network has enough resources to support the user requirement, the neural network based rate control algorithm guarantees the uniform ultimate boundedness of the achieved round trip with respect to an arbitrarily small neighborhood of the desired round trip. In the unfortunate case of high congestion, the controller reduces the rate to the point of congestion escaping. To improve the achievable sending time, introducing the notion of communication channels, whose number is regulated via an appropriately designed algorithm, parallelizes the transmission process. We assume the nonexistence of uncontrollable traffic flows (e.g., UDP) as well as negligible propagation delays. Simulation studies illustrate the approach, highlighting various performance measures (i.e. fairness, buffer level, router utilization) and verifying theoretical analysis.