Reinforcement learning framework for just-in-time material ordering in urban construction: Development, evaluation, and real-time deployment

Efficient material logistics is a persistent challenge in urban construction, where space limitations, delivery uncertainties, and fluctuating demand frequently disrupt site operations. This study presents a Deep Q-Network (DQN)-based decision-support system for optimizing just-in-time (JIT) ordering of steel reinforcement on congested sites. A custom simulation environment was developed using real project data and domain knowledge, capturing realistic constraints such as truck batching, stochastic demand drivers (e.g., weather, crane breakdowns, worker productivity, and delivery punctuality), and short-term planning horizons. The model is trained to minimize total logistics cost, incorporating penalties for storage, shortages, and inefficient deliveries. Performance is benchmarked against fixed, threshold-based, and myopic policies under varying uncertainty conditions. Results show that the DQN outperforms traditional methods in cost efficiency and reliability, particularly under high variability. The trained model is deployed through a user-facing web application built in Streamlit, enabling site managers to input live site conditions and receive interpretable order recommendations. The research advances construction engineering and management information technologies by integrating reinforcement learning (RL) with construction-specific logistics modeling and deploying the resulting policy through a practitioner-oriented digital decision-support interface. This work bridges the gap between algorithmic RL and real-world construction practices, offering a scalable and practical tool to support material planning in dynamic site environments.