Selection hyper-heuristics for healthcare schedulingTools Banerjea-Brodeur, Monica (2013) Selection hyper-heuristics for healthcare scheduling. PhD thesis, University of Nottingham.
AbstractA variety of approaches have been used to solve a variety of combinatorial optimisation problems. Many of those approaches are tailored to the particular problem being addressed. Recently, there has been a growing number of studies towards providing more general search methodologies than currently exist which are applicable to different problem domains without requiring any algorithmic modification. Hyper-heuristics represent a class of such general methodologies which are capable of automating the design of search process via generating new heuristics and/or mixing existing heuristics to solve hard computational problems. This study focuses on the design of selection hyper-heuristics which attempt to improve an initially created solution iteratively through heuristic selection and move acceptance processes and their application to the real-world healthcare scheduling problems, particularly, nurse rostering and surgery admission planning. One of the top previously proposed general hyper-heuristic methodology was an adaptive hyper-heuristic consisting of many parameters, although their values were either fixed or set during the search process, with a complicated design. This approach ranked the first at an international cross-domain heuristic search challenge among twenty other competitors for solving instances from six different problem domains, including maximum satisfiability, one dimensional bin packing, permutation flow shop, personnel scheduling, travelling salesman, vehicle routing problems. The hyper-heuristics submitted to the competition along with the problem domain implementations can now be considered as the benchmark for hyper-heuristics. This thesis describes two new easy-to-implement selection hyper-heuristics and their variants based on iterated and greedy search strategies. A crucial feature of the proposed hyper-heuristics is that they necessitate setting of less number of parameters when compared to many of the existing approaches. This entails an easier and more efficient implementation, since less time and effort is required for parameter tuning. The empirical results show that our most efficient and effective hyper-heuristic which contains only a single parameter outperforms the top ranking algorithm from the challenge when evaluated across all six problem domains. Moreover, experiments using additional nurse rostering problems which are different than the ones used in the challenge and surgery scheduling problems show that the results found by the proposed hyper-heuristics are very competitive, yielding with the best known solutions in some cases.
Actions (Archive Staff Only)
|
Tools
Tools
