A neural heuristic for access network planning

The greedy randomized adaptive search procedure (GRASP) is a well-known metaheuristic for combinatorial optimization. In this paper, we introduce a GRASP for designing the access network topology of a wide area network (WAN). This problem is NP-hard, and can be modeled as a variant of the steiner problem in graphs. The proposed GRASP employs a random neural network (RNN) model in the local search phase, in order to improve the solutions delivered by the construction phase. Experimental results were obtained on 210 problem instances of different topological characteristics, generated using the problem classes in the SteinLib repository, and with known lower bounds for their optima. The algorithm obtained good results, with low average gaps with respect to the lower bounds in most of the problem classes, and attaining the optimum in 48 cases (more than 22% of the test-set).