Active Queue Management With Model Predictive Control

With the evolvement of Internet and the increasing of network applications,the network congestion has become more serious and complex. The goal of con-gestion control is to make use of the network resources e?ectively and fairly, andto improve the network integrated performance and quality of service.Active queuemanagement(AQM) gets more and more recognitions in recent years, which com-bines with TCP end-to-end congestion control being the main approach to solvethe problem of Internet congestion control at present. The research of active queuemanagement is all the while active field in communication group. Control theorieshave been applied to the research of active queue management,the quite represen-tative algorithms are PI/PID algorithms, but the problems about time delay andadaptability can't be e?ectively solved, which encourage us to design new algo-rithms.We expect that the new algorithms combined with TCP congestion controlmechanism could strengthen congestion control mechanisms to support the QoSsystem better.To Internet system in which large delay, complex change and bad distur-bance exist, we exploringly applied model predictive control theories to the re-search of active queue management, designed a new active queue management al-gorithm(MCAQ). Our purpose is to solve the network controlling problems basedon the basic character of model predictive control by combining network with modelpredictive control. The main contents and contributions of this paper are list below:1. We research the rationality of applying model predictive control to networkcontrol in detail. Based on the actuality of applying control theories to networkcongestion control, we summarized some control problems lying in Internet caused by large delay and complex change, for example,compensating delays,adaptability,anti-jamming capability.we analyzed the excellence of the character of model pre-dictive control including predictive model, moving horizon strategy,feedback controlin the process of applying them to network control, in the end we clarified rational-ity and applied value of applying model predictive control to the research of activequeue management algorithm.2. We adopted more exact TCP/AQM nonlinear model. Based on TCP/AQMmodel adopted by most literature about network control, we didn't suppose dropprobability close to zero and ignore time delay anymore, adopted more exactmodel. Simultaneously we derived its nonlinear di?erential function at equilibriumpoint.For applying dynamic matrix control theory to design controller, we evenanalyzed the degree of linearity of specific network model detailedly, presented thebasis for getting the step response coe?cient.3. In the process of designing controller applied DMC theory with constrains,first of all, we considered real constrains of network system explicitly, transformedthe restriction of the maximal queue length of bu?er saving and the maximal dropprobability supporting QoS to output and control constrains of control system, ex-plicit constrains avoid the process of validating controller finally.Then, we describedthe goal of congestion control by optimization functions. Because the problems inpredictive control are transformed to optimization problems finally, in congestioncontrol the goals of queue length closing to expected value possibly and voiding thenetwork frequent shocks caused by dropping packets are described by optimizationfunctions. In the end, we initialized parameters including control horizon, pre-diction horizon, weight matrixes rationally, obtained control quantum of MCAQalgorithm which is operated on network system by solving quadratic programming(QP) problem. 4. By a series of simulation through Matlab, we analyzed the advantages anddisadvantages of MCAQ algorithm. In the basic stimulation , the system didn'texist interference, we proof of MCAQ algorithm correctness, summarized its basicadvantages include that the instantaneous queue length arrived expected value sta-bly and ffeetly, it need shorter adjustment times,and the change of drop probabilityis small. For investigating the ability of MCAQ algorithm against interference, tothree group unmeasurable disturbances which are different in the continual actiontime we did stimulation experiments, the results indicated that MCAQ algorithmpossessed preferable anti-jamming ability against instantaneous interference. Oth-erwise we analyzed robust and compensating delays through stimulations, the con-clusion was that MCAQ algorithm was robust to changing slowly network, andcontrolled system in time due to compensating delays.Although this paper proofed that applying predictive control to network con-gestion control is effective, the research is still at an initial stage, a great numberof problems are unsolved so far. Our works need further studying and perfecting,such as followings: (1)Apply nonlinear predictive control to network congestion con-trol. (2)Compensating dynamic time delay in the process of designing controller.(3)Combining NS2 with Matlab in stimulation through embedding controlled ob-ject into Matlab and controller into NS2.