Simultaneous wide-area and local-access network design
Abstract
Optimal design of data and telecommunications networks involves both synthesis of a high capacity wide-area network and provision for local-access network connecting individual terminals to nodes of the wide-area system. Although the allocation of terminals to local-access networks for different wide-area nodes and structuring of the wide-area network are obviously inter-related, most research to date has treated the problems separately. That is, wide-area networks are designed assuming the assignment of terminals is fixed, and local-access networks are structured without regard to the wide-area. This thesis develops a comprehensive approach to solving the simultaneous wide-area and local-access network design problem for the purpose of reducing overall network costs. The solution determines at least cost the links to be included in the wide-area network and the routes for data messages. We first employ a standard piecewise linear approximation to the nonlinear integer wide-area network design problem with an initial fixed local-access network. We develop a route sparsification scheme to choose a "good" set of routes in order that a primal heuristic can produce improving results. Upon solving the LP we employ a single cluster heuristic to create new groups of terminal nodes by utilizing the dual variables from the LP solution. The dual variables provide the clustering heuristic with information concerning the behavior of the wide-area network. Dual prices help facilitate the construction of new clusters for the purpose of improving solution quality. We then develop a column generation procedure in which clusters become columns and each iteration of the procedure produces new dual prices for constructing new clusters. The procedure continues until no further significant solution improvement. All three aspects of the research--linearization, clustering and column generation--are extensively tested on randomly generated instances. Results show that each of the three can contribute to a better design. Wide-area results from the LP model and primal heuristic out perform those of more standard Lagrangian relaxations in almost all cases. A single stage of clustering produces improvements for most instances, and full column generation leads to even better solutions.
Degree
Ph.D.
Advisors
Rardin, Purdue University.
Subject Area
Operations research|Industrial engineering|Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.