Abstract
This paper proposes a novel hierarchical deep reinforcement learning framework for multi-objective optimization of dynamic logistics scheduling problems. Traditional optimization methods and conventional reinforcement learning approaches struggle with the complexity, uncertainty, and competing objectives inherent in modern logistics operations. Our approach integrates temporal abstraction through a two-level architecture comprising high-level policy networks for strategic planning and low-level execution networks for tactical implementation. We design a Pareto-optimal reward mechanism with adaptive weighting to balance time efficiency, cost control, and service quality objectives while addressing the challenges of sparse rewards and sample efficiency. Extensive experiments on diverse logistics datasets demonstrate the superiority of our approach, achieving 18.4% improvement in service quality, 15.2% reduction in order fulfillment time, and 7.8% decrease in operational costs compared to state-of-the-art methods. Visual analyses and case studies illustrate the framework's ability to generate high-quality Pareto-optimal solutions that effectively balance competing objectives. The proposed framework exhibits particular advantages in dynamic environments with real-time disruptions, maintaining strategic coherence while enabling tactical adaptability with near real-time decision-making capabilities.