Research Report: GPU-based Approaches for Hybrid Metaheuristics

In combinatorial optimization, near-optimal algorithms such as metaheuristics allow to iteratively solve in a reasonable time NP-hard complex problems. Two main categories of metaheuristics are distinguished: population-based metaheuristics (P-metaheuristics) and solution-based metaheuristics (S-metaheuristics). P-metaheuristics are population-oriented as they manage a whole population of solutions, what confers them a good exploration power. Indeed, they allow to explore a large number of promising regions in the search space. On the contrary, S-metaheuristics such as local search algorithms work with a single solution which is iteratively improved by exploring its neighborhood in the solution space. Therefore, they are characterized by better local intensification capabilities. Theoretical and experimental studies have shown that the hybridization between these two classes of metaheuristics improves the quality of provided solutions and the robustness of the metaheuristics [1]. Nevertheless, as it is generally CPU time-consuming it is not often fully exploited in practice. Indeed, experiments with hybrid metaheuristics are often stopped without convergence being reached. That is the reason why, in designing hybrid metaheuristics, there is often a compromise between the number of solutions to use and the computational complexity to explore it. To deal with such issues, only the use of parallelism allows to design efficient hybrid metaheuristics. Recently, graphics processing units (GPU) have emerged as a new popular support for massively parallel computing [2], [3]. Such resources supply a great computing power, are energyefficient, and unlike grids, they are highly available everywhere: laptops, desktops, clusters, etc. During many years, the use of GPU computing was dedicated to graphics and video applications. Its utilization has recently been extended to other application domains [4], [5] (e.g. scientific computing) thanks to the publication of the CUDA (Compute Unified Device Architecture) development toolkit that allows GPU programming in a C-like language [6]. With the emergence of standard programming languages on GPU and the arrival of compilers for these languages, combinatorial optimization on GPU has generated a growing interest. Historically, due to their embarrassingly parallel nature, P-metaheuristics such as evolutionary algorithms have been the first subject of parallelization on GPU architectures. Hence, previous approaches and implementations have been proposed for genetic algorithms [7], [8], particle swarm optimization [9], [10], ant colonies [11], [12], genetic programming [13], [14] and other evolutionary computation techniques [15], [16]. In comparison with previous works on population-based metaheuristics, the spread of solutionbased metaheuristics on GPU does not occur at the same pace. Indeed, the parallelization on