Deep reinforcement learning for optimal resource allocation (II)

Abstract

As vending machines become increasingly intelligent, enabling real-time updates of stock levels, the manual task of restocking remains a logistical challenge. This paper addresses the efficient restocking of vending machines via the Capacitated Vehicle Routing Problem (CVRP), focusing on optimizing routes for limited-capacity vehicles to meet demand without exceeding capacity, while minimizing costs. Traditional heuristics such as LKH-3 have shown robust performance in CVRP but face limitations in scalability and adaptability. This study compares two advanced learning-based approaches—L2D, employing deep reinforcement learning, and NCO, with its innovative light encoder and heavy decoder architecture—against the LKH-3 algorithm. Through detailed experimentation, we evaluate their scalability, computational efficiency, and solution quality.

Our findings reveal that while L2D and NCO exhibit superior generalization capabilities and demonstrate promising scalability to large-scale problem instances, nuances in performance and efficiency metrics highlight their respective strengths and areas for improvement. The comparative analysis not only underscores the potential of learning-based models in overcoming the limitations of traditional heuristics but also delineates the path for future research in integrating the computational intelligence of machine learning with the intuitive problem-solving prowess of heuristic algorithms. This synthesis aims to pave the way for innovative solutions to CVRP and other combinatorial optimization challenges, marking a significant stride toward leveraging artificial intelligence in operational research.