This paper introduces the Generative Flow Ant Colony Sampler (GFACS), a neural-guided probabilistic search algorithm for solving combinatorial optimization (CO). GFACS integrates generative flow networks (GFlowNets), an emerging amortized inference method, with ant colony optimization (ACO), a promising probabilistic search algorithm. Specifically, we use GFlowNets to learn a constructive policy in combinatorial spaces for enhancing ACO by providing an informed prior distribution over decision variables conditioned on input graph instances. Furthermore, we introduce a novel off-policy training algorithm for scaling conditional GFlowNets into large-scale combinatorial spaces by leveraging local search and shared energy normalization. Our experimental results demonstrate that GFACS outperforms baseline ACO algorithms in seven CO tasks and is competitive with problem-specific heuristics for vehicle routing problems.

翻译：本文提出了一种用于解决组合优化（CO）问题的神经引导概率搜索算法——生成流蚁群采样器（GFACS）。GFACS将新兴的摊销推理方法生成流网络（GFlowNets）与前景广阔的概率搜索算法蚁群优化（ACO）相结合。具体而言，我们利用GFlowNets在组合空间中学习构造性策略，通过提供以输入图实例为条件的决策变量先验分布来增强ACO。此外，我们提出了一种新颖的离策略训练算法，通过利用局部搜索和共享能量归一化技术，将条件GFlowNets扩展到大规模组合空间。实验结果表明，GFACS在七项组合优化任务中优于基线ACO算法，并在车辆路径问题上与特定问题启发式算法具有竞争力。