Supply chain management has become increasingly critical as transportation costs rise and urban delivery systems place additional pressure on already congested networks. Last-mile delivery—the movement of goods from distribution centers to end users—represents the most expensive and operationally complex segment of the logistics process, particularly in underdeveloped megacities characterized by unpredictable traffic, limited infrastructure, and sparse real-time data. This paper proposes an integrated framework for optimizing last-mile delivery in such environments, combining social media-based traffic intelligence, machine learning, and deep reinforcement learning. The framework leverages unstructured social media data to enhance traffic prediction and incorporates these predictions into a reinforcement learning model that formulates the delivery problem as a Markov Decision Process. This approach enables dynamic routing decisions that adapt to evolving traffic conditions, with the objective of reducing delivery time, operational cost, and fuel consumption. The framework is evaluated through a case study of Bogotá, Colombia, demonstrating improved routing efficiency and reduced impact of urban traffic uncertainty. The results highlight the potential of integrating alternative data sources with adaptive learning to enhance last-mile delivery performance in data-scarce urban environments.

Keywords

Last-Mile Delivery, Reinforcement Learning, GCN-LSTM, Social Media Analytics, Traffic Prediction, Discrete-Event Simulation