An accurate model of an ATM communication network must take into account the buffer delays which are time-variant, and nonlinearities arising from buffer and source rate saturation and quantization. The time-variant buffer delays in particular play a crucial role in the dynamic behavior of high-speed networks. The ensuing nonlinear time-variant model however is extremely difficult to analyze. All existing work therefore rely on simplified linear piecewise time-invariant models. At simulation stage, network flow controller designs based on such models may appear satisfactory; however, when applied to a high-speed system, the true network often exhibits undesirable behavior in contrast to what is predicted.
In this proposal, a more realistic model for high-speed networks that includes time-variant delays and nonlinearities will be developed. A distributed sensor network (DSN) that operates in a highly dynamic environment is a particularly demanding application where the network time-variance cannot be ignored. Indeed, treating such a network by conventional piecewise time-invariant methods can lead to unacceptable oscillatory behavior.
Communication and processing of signals from a DSN that utilizes a variety of sensors which may be distributed logically, spatially, and geographically, is an essential problem encountered in several areas, such as, robotics, automation, aerospace, medical imaging, etc. In such a DSN, decisions are made at each node in a particular hierarchical level by fusing data communicated from the lower levels. In a highly dynamic environment where the observed scene can change frequently and dramatically, it is essential that importance measures or weights are attached to nodes in the network. This allows allocation of `importance proportional rates' to each node. Such a paradigm, when implemented via an ATM network with the available bit-rate (ABR) option, is shown to allow effective utilization of bandwidth, possibility of capturing transient network dynamics with minimal data loss, and adaptability to fast changes and faults. To accommodate `bursty' traffic, buffer set points are allowed to be varied according to importance of corresponding nodes. Furthermore, network configuration guidelines that allow maximum utilization of the proposed variable bit-rate scheme are provided.
To ensure fault tolerance, each node is provided with a local controller whose parameters are chosen to guarantee the node rate to conform to the allocated weight within a specified settling time. The fundamental requirement of stability of such a time-variant network is achieved via recent results on robust stability. The same methods will be used to provide bounds on maximum delays involved and buffer lengths. This allows effective utilization of the proposed strategy of variable buffer set points. For the assignment of importance measures and weights to each node, Dempster-Shafer evidential reasoning techniques are invoked.
