Moving Horizon Control For Active Queue Management

Internet congestion occurs when the aggregate demand for a resource exceeds the available capacity of the resource. With the rapid development of the Internet, network congestion is occurring more frequently. The network is shared by many competing users including the aggressive flows. Transmission control protocol (TCP) is widely used be-cause of the success of its congestion control mechanism. In the process of network data transmission, TCP congestion control is based on the principle of conservation of packets. However, streaming media applications are usually based on User data protocol (UDP) which provides no end-to-end congestion control. In the absence of end-to-end congestion control, the unresponsive UDP flows can expand to use the vacant bandwidth and bring the network to a congestion collapse. In order to cope with the network congestion, some new congestion control mechanisms are required at routers. The traditional queue man-agement technique at a router is tail drop. When the buffer has become full, the data packets are dropped as the congestion signals to sources. A full queue phenomenon may persist for a long time. Therefore, it is necessary for a router to drop packets before a queue becomes full to avoid congestion. Signaling the congestion information by drop-ping or marking packets at routers is called active queue management (AQM). AQM has become an attractive control strategy for congestion avoidance. Its goal is to maintain a small queue length with enough buffer capacity to absorb the burst data traffic. The random early detection (RED) method, popular at the beginning of studying AQM con-trollers, however, is too sensitive to parameter configuration. A deep insight of the system dynamics is helpful to find new control algorithms. To this end, a fluid-based dynamic model of the TCP was developed by using stochastic theory. Based on this TCP model, the fundamentals of control theory have been used to analyze and develop new AQM schemes. Recently, there is a growing interest in the moving horizon control problem for AQM, which has both theoretical and engineering application significance.The objective of this paper is to present systematic moving horizon control scheme for AQM. Firstly, we propose model predictive control (MPC) based on the extended model of network system as a novel AQM algorithm. The objective is to stabilize the queue length as a target value for improving the network performance and smoothing out the burst traffic under the arbitrary feedback delays. In this scheme, the queue length is predicted based on the extended TCP/AQM system model and the state estimator. Ac-cording to the predicted future queue length in the data buffer, early packets at the router are dropped reasonably by the proposed controller so that the queue length reaches the desired value with minimal tracking error. The drop probability is obtained by optimizing the network performance. Further, randomized algorithms are applied to analyze the ro-bustness of the closed-loop system successfully, and also to provide the stability domain of systems with uncertain network parameters. The performances of the designed algorithm are evaluated through a series of simulations in NS2. The simulation results show that the designed algorithm outperforms RED, PI, and SPI-RED algorithms in terms of stability, disturbance rejection, and robustness.With the large time delay, the order of the extended TCP/AQM system model is increased. In order to cope with this problem, an improved AQM algorithm based on MPC is proposed. Considering the causality of time-delay system, the predicted output is defined again. Then, the control requirement is converted to the optimal control objective, and the drop probability is obtained by solving the optimal problem. Furthermore, the delay-independent stability condition of closed-loop system is derived, which gives the guidelines how to select the parameters of the improved algorithm to assure the stability and fast convergence of the network system. The performances of proposed algorithm are evaluated through a series of simulations in NS2 under the single/multiple topology. The simulation results show that the queue length with designed algorithm reaches the desired value with minimal tracking error and lower drop probability. Moreover, the order of the optimal control problem needs not to increase with the larger time delay in the dynamic network.Considering the large time-delay, complex variations and detrimental disturbances, a novel AQM algorithm based on the constrained MPC for the Internet system is designed. The packet-loss rate is obtained by solving the optimization problem online to improve the adaptive ability of the AQM algorithm. The proposed algorithm adapts to the varying network environment and improves the robustness by moving the optimization horizon, and handles network constraints in the process for determining the packet-loss rate. The step response model and the network model with input delay are used to design the constrained model predictive controller respectively. Then, all of them are converted to quadratic programming (QP) problem. By extending the function of the network simulator NS2, the QP problem can be solved in NS2. The simulation results show the the effectiveness of the designed controller based on the network model with input delay. Comparing the different predictive models, the computational burden of solving QP problem online is described in detail.Finally, building on the T-S fuzzy models, a moving horizon H∞control scheme for nonlinear systems is proposed to guarantee a disturbance attenuation performance for network system with disturbances. Our main contribution is the combination of T-S fuzzy models and the moving horizon H∞control scheme to address the disturbance attenuation issue of nonlinear constrained system. A sufficient condition on the amplitude of the disturbances is given to guarantee the feasibility of the optimization problem at each time. A parameter dependent state feedback control law is adopted, and the corresponding optimization problem is reduced to a convex optimization problem involving linear matrix inequalities (LMIs). The H∞attenuation index of nonlinear moving horizon H∞control problem is adapted online to satisfy time-domain constraints. Finally, the effectiveness of the proposed scheme is successfully demonstrated in the control of AQM at routers.The design processes and the analysis of mentioned AQM algorithms are presented in detail in this thesis. While the comparison and verification of the different predictive models and the guiding principles for the controller parameters are given. Moreover, in order to validate the efficiencies and cost of the proposed approaches, we give simulation results for each approach, which is discussed in detail from predictive model choosing, controller parameter regulating, simulation environment testing. The simulation results indicate that the control effects with the proposed approaches are satisfactory and their own features are explicit.Deeper research work needs to be done since some problems are still remain to be solved, such as how to speed up the online calculation of QP problem by improving NS2, or how to reduce the conservativeness of the T-S fuzzy models based moving horizon H∞control scheme etc. Moreover, when the AQM algorithm based on unconstrained MPC can be embed into the real network, more valuable conclusion will be obtained.