Transformer-Based Inventory Decision for Multi-Location Newsvendor Systems under Limited Data

In classical inventory models, decisions are optimized based on a well-specified stochastic model that characterizes the real system, which yields clear structure and theoretical insights. This approach, however, often comes at the cost of flexibility and tractability. Recent advancement in artificial intelligence (AI) offers a compelling alternative, but its reliance on large datasets, black-box nature, and lack of theoretical guarantees limit its adoption in practice.*Project Description*The project aims to bridge the classical and the AI paradigms by proposing a transformer-based inventory policy that embeds inventory-specific domain knowledge nto generative AI, which enables efficient decision-making with limited data. Our approach is threefold:We design a transformer architecture tailored for inventory decision-making based on historical observations that allows efficient optimization. The transformer takes historical demand-covariate pairs as inputs and automatically generates an inventory decision for the next period in real time.We embed inventory-specific domain knowledge into the transformer decision by training the transformer model on simulated demand trajectories drawn from a series of demand models. The demand trajectories are properly augmented through certain transformations, which align the learned policies with the newsvendor solution structure known from classical inventory theory.For multi-location inventory decision, we exploit the cross-attention mechanism to pool information across sites, which enables cross learning between different locations.We further incorporate capacity constraints through Lagrangian relaxation. Our model can thus make efficient real-time decisions for large-scale newsvendor systems faster than conventional methods.*Preliminary Results and Significance*We have developed a preliminary transformer model and an algorithm for the single-location newsvendor problem. Under a linear demand model, we can show that our approach leads to a uniformly optimal solution among a broad class of decision rules developed in the literature. The project will result in a scalable tool that automatesdata-driven decision making for large inventory systems, and accelerate the adoption of AI-driven inventory optimization in real-world operations. By bridging classical operations management (OM) insights with modern generative AI, this project advances both theory and practice in data-driven inventory management.