本研究針對多目標最短路徑問題(Multi-Objective Shortest Path， MOSP) 面臨的 標籤爆炸與即時性瓶頸，提出一種結合圖神經網路(GNN) 與經典解碼的混合尋路架 構。我們設計了基於查詢條件化的GNN 模型，將道路幾何特徵、起終點標記與動態目 標偏好向量λ 融合，使其具備動態調整偏好的圖卷積表徵能力。數據生成階段，採用 加權Dijkstra 模仿學習高效生成70，000 筆偏好導引的訓練樣本；解碼階段，設計幾何 引導式束搜尋解碼器，結合距離衰減啟發式估計以克服繞路異常並加速邊剪枝。 實驗結果顯示，在真實舊金山路網數據上，所提架構成功實現了動態偏好路線權 衡與加速搜尋。消融實驗選定之Epoch 30 模型在短、中、長距離場景下，單次查詢執 行時間介於33.1 至110.4 毫秒，中距離較經典Martins 演算法加速達22.0 倍。雖然在 單次偏好查詢的搜尋速度上仍慢於精確演算法BOA*，但本模型無需針對特定終點進 行兩次反向全局預處理，且具備GPU 併行批次推論的吞吐量優勢。本研究證實，將條 件式圖表徵學習與啟發式解碼結合，是求解大規模多目標導航中兼顧動態偏好與可擴 展性的一種可行且具啟發性的方案。;This study presents a hybrid routing framework that combines Graph Neural Networks (GNNs) with a heuristic decoder to address the label explosion and real-time execution challenges in the Multi-Objective Shortest Path (MOSP) problem. We design a Query-Conditioned GNN that integrates road geometric features， origin-destination indicators， and a dynamic preference vector λ， enabling GNN layers to adjust representations based on varying target preferences. For sample generation， we employ a scalarized Dijkstra imitation learning pipeline to efficiently generate 70，000 preference-guided training samples. During decoding， we propose a geometrically-guided beam search decoder integrated with target distance decay estimates to prune low-confidence branches and overcome detour anomalies. Experimental results on the San Francisco road network demonstrate the effectiveness of the proposed framework in enabling dynamic route trade-offs and accelerating search. Using weights selected at Epoch 30， the GNN model achieves single-query execution times ranging from 33.1 to 110.4 ms across short， medium， and long-distance scenarios， achieving a 22.0-fold speedup compared to Martins’ algorithm for mediumdistance queries. Although the exact solver BOA* is faster in single-preference queries， our model requires zero target-dependent backward preprocessing and benefits from GPUaccelerated batch inference. This work confirms that combining query-conditioned graph representation learning with heuristic decoding is a viable and promising approach for large-scale MOSP navigation.
