Lookahead Policy and Genetic Algorithm for Solving Nurse Rostering Problems

Previous research has shown that value function approximation in dynamic programming does not perform too well when tackling difficult combinatorial optimisation problems such as multi-stage nurse rostering. This is because the large action space that needs to be explored. This paper proposes to replace the value function approximation with a genetic algorithm in order to generate solutions for the dynamic programming stages. Then, the paper proposes a hybrid approach that generates sets of weekly rosters with a genetic algorithm for consideration by the lookahead procedure that assembles a solution for the whole planning horizon of several weeks. Results indicate that this hybrid between a genetic algorithm and the lookahead policy mechanism from dynamic programming exhibits a more competitive performance than the value function approximation dynamic programming investigated before. Results also show that the proposed algorithm ranks well in respect of several other algorithms applied to the same set of problem instances. The intended contribution of this paper is towards a better understanding of how to successfully apply dynamic programming mechanisms to tackle difficult combinatorial optimisation problems.