Neural network approximation of model predictive controller for congestion control of TCP/AQM networks

Due to the excellent properties of the model predictive controllers (MPC) in implementing on nonlinear and time varying systems, utilizing these controllers as Active Queue Management (AQM) strategy is proposed for congestion control of computer networks. However, high computational demand to solve the optimization problem exist in these controllers is a major obstacle when they are applied on fast large-scale constrained systems such as the computer networks. Small signal linearized model of the nonlinear TCP/AQM network is used to design MPC controller and then a neural network is trained to approximate the model predictive control strategy. Using this approach, due to the parallel processing property of the neural networks, some of the computations can be done in parallel form, therefore, computational effort is reduced compared to the commonly used MPC schemes. The performance of the proposed controller in desired queue tracking and disturbance rejection is compared with two well-known AQM methods, PI and RED.