Topological Design of Distributed Data Communication Networks Using Linear Regression Clustering

A new network decomposition-optimization algorithm is presented and evaluated via computer simulation. The network is first partitioned into subnetworks on the basis of lines fitted via linear regression to the node locations weighted by their traffic loads. Each node is then connected via a data line of appropriate capacity to a concentrator-multiplexer. The concentrators are then regarded as nodes, and the process is repeated as often as required. The resulting singly connected network consists of a hierarchy of concentrators whose number, capacity, location, and interconnection are selected as an inherent part of the design procedure to minimize network cost. The linear regression clustering (LRC) design procedure is evaluated via computer simulation by comparing the costs and performance of the resulting networks with that which results from use of an algorithm based on the generally applicable network design approach of iterative local optimizations or "search" procedures. The data supplied to the LRC and search algorithms include randomly generated node locations and traffic matrices, and specific (realistic) cost vs. capacity schedules for data lines and concentrators. Comparisons of network costs, queuing and transmission delays, network reliability and network design costs show the LRC algorithm to be markedly superior to the search algorithm. The paper includes a brief discussion of our results, and their implications.