Designing state-of-the-art metaheuristics: What have we learned from the parallel-machine scheduling problem with setups?

For over two decades, the unrelated parallel machines scheduling problem with sequence-dependent setup times has engaged the research community, finding relevance in numerous practical applications. Despite the progress of exact approaches, the field of metaheuristics has seen a proliferation of new methods that often claim superiority over existing ones. In this paper, we survey, implement, and test a group of 12 metaheuristics for makespan minimization. An extensive computational study is conducted in an attempt to establish the components that lead to better results. We discuss common design choices, solution representation schemes, and acceleration procedures to reduce computational effort. Experimental testing on 3260 test instances from two well-known benchmarks shows that a simple simulated annealing algorithm with few control parameters, complete solution representation, and diverse neighborhood structures present the best results under a standardized set of experimental conditions. All source codes are available to the scientific community in a reproducible capsule.