Machine Learning for Sequential Decisions in Logistics

Logistics decision-making involves complex, sequential processes influenced by uncertainty and real-time constraints. Traditional operations research (OR) methods provide structured approaches, but machine learning (ML) offers additional flexibility in adapting to dynamic environments. This dissertation examines how ML can support decision-making in logistics across three areas: supply logistics, distribution logistics, and revenue management.
In supply logistics, ML-based models address inventory management challenges. Chapter 2 introduces a hybrid ML-OR approach for replenishment and inspection decisions, integrating neural networks with optimization. Chapter 3 applies reinforcement learning (RL) to dual sourcing, improving supply chain resilience under demand uncertainty. Chapter 4 develops a neighborhood search algorithm combined with RL for large-scale inventory control problems.
In distribution logistics, ML enhances routing and customer selection. Chapter 5 uses predictive models to estimate transportation costs for multi-period customer selection. Chapter 6 applies RL to dynamic vehicle routing problems (DVRPs), enabling real-time adaptation to demand fluctuations.
In revenue management, ML-driven models dynamically adjust logistics service offerings. Chapter 7 explores ML-based pricing and selection for parcel lockers, while Chapters 8 and 9 optimize attended home delivery (AHD) time slots using ML and decision-focused learning approaches.
This research distinguishes between data analytics, which extracts insights from historical data to support decision-making, and decision analytics, which directly optimizes sequential decisions. A structured framework is introduced to guide ML integration in logistics. This framework supports the development of adaptive, data-driven decision-making processes that enhance efficiency, flexibility, and resilience in logistics operations.