A Self-learning Optimization Technique for Topology Design of Computer Networks

摘要

Topology design of computer networks is a constrained optimization problem for which exact solution approaches do not scale well. This paper introduces a self-learning, non-greedy optimization technique for network topology design. It generates new solutions based on the merit of the preceding ones. This is achieved by maintaining a solution library for all the variables. Based on certain heuristics, the library is updated after each set of generated solutions. The algorithm has been applied to a MPLS-based IP network design problem. The network consists of a set of Label Edge Routers (LERs) routing the total traffic through a set of Label Switching Routers (LSRs) and interconnecting links. The design task consists of - 1) assignment of user terminals to LERs; 2) placement of LERs; and 3) selection of the actually installed LSRs and their links, while distributing the traffic over the network. Results show that our techniques attain the optimal solution, as given by GNU solver - lp_solve, effectively with minimum computational burden.