Hopfield neural networks for timetabling: formulations, methods, and comparative results

Abstract

This paper considers the use of discrete Hopfield neural networks for solving school timetabling problems. Two alternative formulations are provided for the problem: a standard Hopfield–Tank approach, and a more compact formulation which allows the Hopfield network to be competitive with swapping heuristics. It is demonstrated how these formulations can lead to different results. The Hopfield network dynamics are also modified to allow it to be competitive with other metaheuristics by incorporating controlled stochasticities. These modifications do not complicate the algorithm, making it possible to implement our Hopfield network in hardware. The neural network results are evaluated on benchmark data sets and are compared with results obtained using greedy search, simulated annealing and tabu search.