Control Seminar

Stochastic Observability and Network Congestion Control

Robert BitmeadPofessorUniversity of California, San Diego

A Control Theorist, the speaker – let's call him CT – studies TCP/IP congestion control as if it were a feedback control system. This leads to the reinterpretation of the pieces of the network from source-bottleneck node-destination to controller-plant-measurement and the posing of a question concerned with state observability of the bottleneck node by the source node using the input and acknowledgement sequence measurements. The whole thing is described by a Hidden Markov Model. CT is surprised by the dearth of results concerning nonlinear stochastic system observability and wonders whether this will prevent him deriving a better feedback approach to congestion control. CT will explain an approach to such a system property. As is clear after a moment's thought, nonlinear observability depends on the feedback control law operating. Using the new measure of stochastic observability, CT will show that the common AIMD rate control law actually has the effect of making the bottleneck node's state strongly observable. This will lead to the revelation that existing TCP/IP congestion control laws actually implement a dual adaptive control solution. This was a big surprise for CT. Perhaps it will be for you too.

Professor Robert Bitmead joined the Jacobs School's Mechanical and Aerospace Engineering Department in 1999 after more than 16 years at the Department of Systems Engineering at the Australian National University in Canberra, Australia. Eight of those 16 years were spent as the Executive Director of the Cooperative Research Centre for Robust and Adaptive Systems, an incorporated joint venture between the University, government labs and industry. Professor Bitmead is a Fellow of the IEEE.

Sponsored by

Eaton, Ford, GM, Toyota and the MathWorks