EVENT DOSSIER
The researchers successfully solved the Vehicle Routing Problem with Time Windows (CVRPTW) using Graph Neural Networks (GNN)-guided graph coarsening techniques and adaptive QUBO penalty methods on the D-Wave Advantage2 quantum annealing machine. This method replaced traditional manual tuning with unified GNN configuration, achieving 100% feasibility for all families at N=10 in the Solomon benchmark test, and increasing the feasibility of Type R samples from 80% to 100%. At larger scales (N=80, 100), its performance was significantly better than that of traditional heuristic algorithms. Additionally, adaptive penalty calibration reduced the violation mean from 33.0 to 0.06, and the rate of feasible samples in hardware experiments increased from 0.02% to 39%, verifying the conditional effect. The generated QUBO size was approximately 5-6 times larger than that of traditional methods.
研究人员在 D-Wave Advantage2 量子退火器上利用图神经网络（GNN）引导的图粗化与自适应 QUBO 惩罚，解决了带时间窗的车辆路径问题（CVRPTW）。针对 Solomon 基准测试，该方法通过 GNN 统一配置替代手工调优合并评分，在 N=10 时实现所有 Solomon 族 100% 可行性（R 型从 80% 提升至 100%），且在 N=80、100 时显著优于传统启发式算法。同时，通过自适应惩罚校准及约束处理，将原始约束违规均值从 33.0 降至 0.06，并验证了硬件实验中的条件效应，使可行样本率从 0.02% 增至 39%，且生成的 QUBO 规模约为传统方法的 5-6 倍。