Neuro Fuzzy Model Predictive Control of AQM Networks Supporting TCP Flows A. R. Maghsoudlou 1 , R. Barzamini 2 , S. Soleimanpour 2 , J. Jouzdani 2 1 Azad Islamic University-Aliabad Katoul Branch 2 Amirkabir University of Technology (Tehran Polytechnic) afshin_m_ir@yahoo.com 2 , Barzamini@aut.ac.ir 2 , S.Soleimanpour@aut.ac.ir 2 , Javid.Jouzdani@yahoo.com 2 Abstract – One of the challenges in designing computer networks is "queue management and congestion avoidance". There are several studies for congestion reduction and controlling such as Random Early Detection (RED) and its variants. More recent works on developing congestion avoidance methods include modeling a TCP flow in an Active Queue Management (AQM) of a bottlenecked network link. Rather than classical control theories, that are applied to improve performance and stability of network flows, some studies are developed based on new control tools such as neural networks. In this article a neuro-fuzzy controller for Active Queue Management is proposed. In this model, the number of neurons are determined according to complexity of the model and instead of using random values for initializing network’s weights, near to optional values are used. The proposed method decreases the network training time and ensures convergence with higher probability. The results of this method show superior performance over other previous controllers such as classical and neural networks methods. I INTRODUCTION Congestion avoidance and congestion control are debatable issues in developing data communication networks and have received a great attention in recent years. A feedback based congestion control scheme is required to ensure trouble free data transfer in a bottlenecked network link. Also prediction of queue situation in a bottleneck node is a very useful tool for avoiding congestion in a node. On the other hand, an efficient use of existing resources prevents data loss that causes retransmission and falling in a wrong loop. The existing data transmission control protocol (TCP) is a window based protocol. Transmitter window size in this protocol is a controllable item based on feedback from the receiver or network routers and hence the transmission rate of data source can be controlled. For instance, in TCP Reno, packet loss is used as a feedback signal for reducing window size in the source while window size in transmitter increases gradually before this congestion notification from the network. One of the most popular congestion avoidance schemes that recently has been proposed for TCP is Random Early Detection [1] and its variants FRED[2], SRED[3] which controls window size based on the queue length in bottleneck. RED marks incoming data packets with probability proportional to the average queuing length and once the source receives ACK of a marked packet, it adjusts its window size[1]. The input –output description of RED is shown below in Figure 1. Figure 1- RED Input Output Map The input-output relationship shown in Figure 1 can be governed by: (1)                 otherwise q q q q p p q q p q q p min max min max max min 1 0 Although this scheme can avoid packet loss but because of nonlinear nature of TCP behavior, such a controller seems not to be efficient for holding queue length in a fixed size and is not robust against number of TCP connections [5]. Fixed queue length leads to a fixed round trip time and with fixed delay, a more efficient control over the system can be achieved. In this paper, a neural predictive control is employed to learn the probability of the packet marking. It is shown that the adaptive nature of such a controller make it possible to get a fixed queue length and achieve a robust performance against changing in the number of TCP connections. The rest of the paper is organized as follows. In section II, a new model of TCP connection based on TCP protocol is proposed that does not have the shortcoming of existing models. In section III, based on the proposed model, a neuro-fuzzy predictive controller is introduced. In section IV, simulation results are presented and a comparison is made to the results obtained by classical controllers such as those presented in [5]. II TCP MODELING A Window Size Dynamics Recent investigations of TCP [11], [12], [13] have been carried out from a source centric point of view. The assumption underlying all these works is that a source sends out the packets to the network with an associated loss Ninth ACIS International Conference on Software Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing 978-0-7695-3263-9/08 $25.00 © 2008 IEEE DOI 10.1109/SNPD.2008.64 226 Ninth ACIS International Conference on Software Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing 978-0-7695-3263-9/08 $25.00 © 2008 IEEE DOI 10.1109/SNPD.2008.64 226