A multi-improvement local search using dataflow and GPU to solve the minimum latency problem

Optimization problems have great importance in the industrial field, specially for supply chain management and transportation of goods. Many of these problems are classified as NP-Hard, thus there is no known algorithm to find their exact (global optimal) solutions in polynomial time. Therefore, fast heuristic strategies are generally employed, specially those with the ability to escape from poor quality local optima, called metaheuristics. In general, the Local Search is the most computationally expensive phase of a metaheuristic, thus requiring a good use of all available computing resources. In this work, we explore state-of-the-art GPU processing local search modules (called neighborhood structures) from literature, together with a proposed Dataflow model with Multiple Output Gates. Although these neighborhood structures are classic for routing problems in literature, these are typically explored in a sequential manner, named Variable Neighborhood Descent. In this work, we demonstrate how to use these neighborhoods on multiple collaborative computing devices, building novel efficient Local Searches for a challenging optimization problem: the Minimum Latency Problem. Finally, we present experiments with the proposed distributed strategies on the Minimum Latency Problem, indicating the gains over previously proposed sequential/parallel approaches in literature, and also the current limitations to deal with larger problem instances. (C) 2020 Elsevier B.V. All rights reserved.