Beam search is a tree search procedure where, at each level of the tree, at most W nodes are kept. This results in a metaheuristic whose solving time is polynomial in W. Popular for single-objective problems, beam search has only received little attention in the context of multi-objective optimization. By introducing the concepts of oracle and filter, we define a paradigm to understand multi-objective beam search algorithms. Its theoretical analysis engenders practical guidelines for the design of these algorithms. The guidelines, suitable for any problem whose variables are integers, are applied to address a bi-objective 0-1 knapsack problem. The solver obtained outperforms the existing non-exact methods from the literature.
Citation
Department of Business Administration, University of Vienna, 2015