Model Predictive Neural Control ofTCP Flow in AQM Network
Kourosh Rahnamai, Kevin Gorman Andrew Gray
Electrical Engineering Department Jet Propulsion Laboratory
Western New England College California Institute ofTechnology
Springfield, M 01119 USA Pasadena, CA 91109
Abstract - Many research papers have been published on first present a [1-3] model for a typical TCP flow, based on
RED (Random Early Detection) and variants of RED [1-12]. the fluid dynamics theories.
Recently many articles have been presented on modeling a
Transmission Control Protocol (TCP) flow in an Active Queue I. TCP MODELING
Management (AQM) of a bottlenecked network link [1-3].
Classical control theories have also been applied to achieve or The following nonlinear equations describe the behavior of a
improve stability of the network flow . In this paper we typical bottlenecked network gateway
present a Neural Network (NN) Model Predictive Control (MPC)
of TCP flows. We show the robust adaptive behavior of the MPC
optimal controller under modeling errors and system dynamic d W(t) 1 W(t) W(t - R(t))
changes. We also show the superior transient and steady state -p(t - R(t)) (1)
behavior as well as general stability of MPC as compared to the dt R(t) 2 R(t - R(t))
Classical PI controller.