On-line gradient algorithms for network control

Gradient algorithms are natural tools for on-line adaptive network control. They have the major advantage of relating directly to measurable cost/performance functions. Difficulties in using such algorithms arise primarily in their computational complexity, in the possibility of effectively distributing the computation accross the network, and in the time-response of the network to the controls. The purpose of this paper is to discuss these issues, and to present analytical and simulation results which point towards the potential benefit of gradient algorithms for future high speed -network management and control.