Personalised urban tourist bus routing via multi-objective deep reinforcement learning: A hybrid PCA-double DQN approach

Conventional routing approaches frequently fail to address the specific operational needs of tourist buses, often overlooking physical vehicle constraints and the heterogeneous preferences of tourists. To address these limitations, this paper presents an application-oriented integrated routing framework by integrating a hybrid PCA-Double DQN approach to address the Multi-Objective Tourist Trip Design Problem (MOTTDP). The methodology integrates Principal Component Analysis (PCA) and K-Means clustering to extract relatively interpretable latent tourist profiles from survey data. These preference weights are then dynamically injected into a Double Deep Q-Network (Double DQN) agent, enabling context-aware routing optimisation under different group characteristics while mitigating the overestimation bias commonly associated with standard Reinforcement Learning (RL) applications in routing. Simulation-based validation conducted using the real-world road network topology of Da Nang, Vietnam, demonstrates promising simulation-based performance of this integrated framework against Dijkstra, Standard DQN, and Proximal Policy Optimisation (PPO) baselines. The proposed model achieves ∼ 91.2 % Semantic Attribute Coverage (Preference Alignment) and reduces fuel consumption by ∼ 25 % compared with the distance-based Dijkstra benchmark. These findings suggest that the framework can provide a promising simulation-based balance to the trade-off between maximising overall tourist preferences and minimising operational costs.