運用強化學習技術規劃幸福巴士路線之研究－以南投市為例

Abstract

中文摘要 本研究針對偏鄉地區交通服務不足的問題，探討以深度強化學習（Deep Q-Network, DQN）為基礎的智慧化公車路線規劃模式，並以南投市為實證場域。由於傳統固定式公車路線在地形破碎、人口分布零散的地區難以有效覆蓋居民需求，導致交通可及性低落、公共運輸資源使用效率不彰。為此，本研究整合票證資料、手機信令資料與問卷調查，建立符合在地需求的路網與節點評分機制，進而應用DQN模型進行幸福巴士路線規劃。 為驗證人工智慧方法於偏鄉公車路線規劃的成效，本研究以涵蓋率、服務高需求點、總路徑長度及重疊比例等指標，比較DQN模型與專家路線的表現。結果顯示，DQN所生成的幸福公車路線成功覆蓋專家規劃重點區域東山里與內興里，涵蓋率分別由44.51% 提升至73.17%，及由 36.82% 提升至 91.33%；同時在路線總長度與彎繞度上略優於人工規劃，並將整體路線與個別路段與既有公車路線的重疊比例分別降至5.11% 與5%。此外，敏感度分析結果亦顯示，獎勵函數中的「重要點位權重」與「重疊既有路線懲罰」為關鍵參數。若移除重要點位權重會犧牲公共運輸可及性，而取消重疊懲罰則會降低資源使用效率。綜合評估可確認，DQN模型在整體路線規劃品質上優於傳統方法設計之專家路線，並具高度調適性與可行性，未來可作為偏鄉地區導入智慧交通系統的重要參考依據。

Abstract This study explores an intelligent bus route planning model based on Deep Q-Network (DQN) to address insufficient transportation services in rural areas, with Nantou City as the case study. Traditional fixed-route buses are often ineffective in fragmented terrains with dispersed populations, resulting in poor accessibility and inefficient utilization of transport resources. To overcome these challenges, the research integrates smart card data, mobile signaling data, and questionnaire surveys to develop a demand-oriented network and node scoring mechanism, which is then applied to the DQN model for Demand Responsive Transit System (DRTS) route planning. The study evaluates DQN-generated routes against expert-designed ones using indicators such as coverage rate, service to high-demand points, total route length, and overlap ratio. Results indicate that the DQN-generated DRTS routes successfully covered the key areas emphasized in expert planning—Dongshan and Neixing Villages—raising coverage rates from 44.51% to 73.17% and from 36.82% to 91.33%, respectively. At the same time, the routes slightly outperformed manual planning in terms of total length and curvature, while reducing the overlap with existing bus services to 5.11% for the overall network and 5% for individual segments. Sensitivity analysis further revealed that “importance weighting of key nodes” and “overlap penalties” are critical parameters, as removing them sacrifices accessibility and resource efficiency. Overall, the findings confirm that the DQN model outperforms expert-designed routes in planning quality, demonstrating strong adaptability and feasibility, and serving as a valuable reference for implementing intelligent transportation systems in rural regions.