Adaptive signal processing
Data networks
Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence
Genetic Algorithms in Search, Optimization and Machine Learning
Genetic Algorithms in Search, Optimization and Machine Learning
| Hi-index | 0.00 |
Routing systems for datagram packet-switched networks' iteratively applying Shortest Path First (SPF) algorithms on load-based link cost metrics exhibit poor stabilization and convergence properties at moderate traffic loads without the addition of experimentally determined Bertsekas Additive Bias Factors to aid in damping undesirable oscillations. Routing systems which iteratively apply SPF Algorithms on load-varying link costs implicitly assume routing assignments can be independently modified without affecting the validity of cost estimates on other links. Even under stationary offered loads, this assumption is less accurate as the load increases due to nonlinear dynamical interactions among traffic flows. Failure to compensate for this independence assumption in iterative SPF computations leads to dynamical instabilities above moderate offered loads. Unfortunately, while Bertsekas Additive Bias Factors improve stability, they also reduce the sensitivity of the routing system to network congestion. As alleviation of traffic congestion is often an important service requirement for a routing system, this is a serious drawback. This paper is motivated by a search for techniques to improve the convergence of routing systems upon near-optimum sets of routing assignments for current traffic conditions, without introducing a loss of congestion sensitivity. Due to the limitations of Markov Chain Models of packet-switching networks for our application, a network simulation is developed which provides a more accurate reflection of network dynamics, as it incorporates bounded queue and minimum interarrival time constraints. Utilizing this simulation, it is demonstrated that a Genetic Algorithm which estimates the near-optimal number of routing assignment changes allowed by the Dijkstra SPF Algorithm during each routing update interval, enables substantial performance advantages for networks operated at or above moderate offered loads. This adaptive feedback compensation technique achieves these performance improvements by avoiding the Kauffman and Eigen Catastrophes, although additional processing and slight increases in routing update packet lengths are required. In a broader context, the technology of adaptive feedback compensation should prove a relevant to a variety of real-time distributed adaptive control applications.