First-order Linear Programming in a Column Generation Based Heuristic Approach to the Nurse Rostering Problem

A heuristic method based on column generation is presented for the nurse rostering problem. The method differs significantly from an exact column generation approach or a branch and price algorithm because it performs an incomplete search which quickly produces good solutions but does not provide valid lower bounds. It is effective on large instances for which it has produced best known solutions on benchmark data instances. Several innovations were required to produce solutions for the largest instances within acceptable computation times. These include using a fast first-order linear programming solver based on the work of Chambolle and Pock to approximately solve the restricted master problem. A low-accuracy but fast, first-order linear programming method is shown to be an effective option for this master problem. The pricing problem is modelled as a resource constrained shortest path problem with a two-phase dynamic programming method. The model requires only two resources. This enables it to be solved efficiently. A commercial integer programming solver is also tested on the instances. The commercial solver was unable to produce solutions on the largest instances whereas the heuristic method was able to. It is also compared against the state-of-the-art, previously published methods on these instances. Analysis of the branching strategy developed is presented to provide further insights. All the source code for the algorithms presented has been made available on-line for reproducibility of results and to assist other researchers.