Network Congestion Control Robust Aqm Algorithm

As network is getting bigger and more complex, especially the wide application of multimedia, the exclusive rely on end host congestion mechanism can't meet the quality of service (QoS). To this end, the network itself must participate in congestion control. Active queue management (AQM) as congestion mechanism at routers has received great attention from researchers for its great effect on the improvement of QoS. The current AQM work suffer from the weakness that its performance is affected by network parameter change. For this reason, this thesis uses robust control and neuron network control theory to design three new AQM algorithms which show great robust property against parameters change and uncertainty. The contributions of this thesis are as follows:(1) A new robust algorithm called RC is proposed based on H-infinity control theory, considering a more realistic link capacity disturbance model, which takes the bandwidth occupied by short-lived connections as TCP/AQM system disturbance different from the general available bandwidth disturbance model. Not only takes account of the effect of delay to the TCP/AQM system, but also use on-line TCP window size estimation to reduce computation requirement by state estimator, the parameter of controller is obtained through the standard LMI problem.(2) Uses state space to describe the TCP network congestion control system as linear system with state and input delay. A state-observer based Active Queue Management controller is designed by H-infinity robust control, fully considering the effect of saturated input and network parameters change. The observer is used to obtain the immeasurable real TCP window size, avoiding the error by approximation. An appropriate Lyapunov functional is chosen to turn the control law problem into the Liner Matrix Inequality (LMI).(3) Analyzes the property of traditional REM scheme and points out the weakness of REM is lacking adequate adaptability to complicated time-varying TCP dynamics. In order to improve its robustness, we fully utilize neuron network's online learning ability to adjust the value of price. Considering the big influence of parameter (?) to overall performance, gradient algorithm is used to tune its value to catch the change of the network environment to further improve its adjustability and robustness.