MULTI-AGENT REINFORCEMENT LEARNING FRAMEWORK FOR AUTONOMOUS TRUCK DISPATCHING WITH LLM-AUGMENTED DECISION SUPPORT IN DYNAMIC LOGISTICS ENVIRONMENTS

Efficient truck dispatching is a fundamental problem in logistics systems, where decisions must be made under dynamic, uncertain, and multi-constraint environments. Traditional rule-based and optimization-based dispatching systems struggle to adapt to real-time changes in demand, traffic conditions, driver availability, and market fluctuations. This paper proposes a hybrid intelligent dispatching framework that integrates Multi-Agent Reinforcement Learning (MARL) with Large Language Model (LLM)-based decision support to enable autonomous and adaptive truck dispatching. The system models dispatching as a sequential decision-making problem in which multiple agents collaborate to optimize global system rewards such as profit maximization, empty-mile reduction, and delivery efficiency. The proposed architecture introduces specialized agents for truck allocation, route optimization, load selection, and risk evaluation, coordinated through a reinforcement learning environment. In addition, an LLM-based reasoning layer interprets unstructured inputs such as broker messages, operational constraints, and natural language dispatch instructions. The framework is evaluated in a simulated logistics environment reflecting real-world uncertainties. The results demonstrate that the proposed hybrid system improves dispatch efficiency, decision consistency, and operational profitability compared to classical rule-based and single-agent optimization approaches.