A hybrid CPU-GPU scatter search for large-sized generalized assignment problems

Abstract

In the Generalized Assignment Problem, tasks must be allocated to machines with limited resources, in order to minimize processing costs. This problem has several industrial applications and often appears as substructure of other combinatorial optimization problems. By harnessing the massive computational power of Graphics Processing Units in a Scatter Search metaheuristic framework, we propose a method that efficiently generates a solution pool using a Tabu list criteria and an Ejection Chain mechanism. Common characteristics are extracted from the pool and solutions are combined by exploring a restricted search space, as a Binary Programming model. Classic instances vary from 100–1600 jobs and 5–80 agents, but due to the big amount of optimal and near-optimal solutions found by our method, we propose novel large-sized instances up to 9000 jobs and 600 agents. Results indicate that the method is competitive with state-of-the-art algorithms in literature.