Multi-Objective Deep Reinforcement Learning for the Optimization of Dynamic Large-Scale Vehicle-Cargo Matching Problem

This study presents a novel deep reinforcement learning algorithm, Deep Q-Learning with Long Short-Term Memory (LSTM) and a Multi-Head Self-Attention Mechanism (DLSAM), for the complex vehiclecargo matching problem in logistics. The problem is formulated as a Markov Decision Process, guiding the selection of optimal vehicles for cargo shipments at each decision stage to meet dynamic and potentially large-scale operational demands. The framework of the proposed algorithm employs Deep Q-Learning to optimize the matching strategy through an action-value function, approximated by a deep neural network. LSTM networks are integrated to capture long-range temporal dependencies, essential for modeling the dynamic characteristics of vehiclecargo matching. Concurrently, a Multi-Head Self-Attention mechanism enables the model to identify and assign varying levels of importance to crucial input features, thereby enhancing matching precision. An encoder-decoder architecture further aids in extracting global features from input data and generating effective decision sequences. Extensive experiments on diverse datasets, including scenarios of varying operational scales, compared this methodology with several algorithms such as Policy Gradient and standard Deep Q-Learning. The results demonstrate the superior performance of the developed algorithm across multiple objectives: type matching, time matching, loading rate, and empty running rate. Notably, this model achieves significantly improves type matching and time matching accuracy, while reducing the rates of empty running and empty loading. These evaluations underscore the algorithm’s robustness, stability, and adaptability for diverse and large-scale vehicle-cargo matching scenarios.