Large-Scale Supply Chain Inventory Optimization Using Recurrent Neural Networks

The rapid advancement of AI, driven largely by Moore's law, has opened new possibilities for large‑scale supply‑chain management, where traditional linear and integer programming methods struggle to exploit modern parallel computing resources.

The authors focus on the inventory optimization problem within massive Bill‑of‑Materials (BOM) networks, exemplified by a project for a leading manufacturer involving over 500,000 nodes and 4 million links, where safety stock must be allocated to a small subset of items.

Conventional models such as stochastic service or guaranteed service are limited by network structure assumptions and become infeasible for cyclic or highly connected BOM graphs, prompting the team to adopt a simulation‑based approach that operates on a daily base‑stock policy.

To overcome the prohibitive computational cost (originally O(T n²) with n nodes), the team reformulated the simulation as tensor operations, representing the BOM adjacency matrix in TensorFlow, applying sparse‑matrix techniques to reduce memory usage, and constructing a custom computation graph that uses back‑propagation (BP) instead of IPA for gradient calculation.

They further introduced L1‑regularization into the objective to promote sparsity of safety‑stock locations, and designed a two‑stage optimization algorithm: a stochastic iterative thresholding method followed by stochastic gradient descent.

Experimental results show speedups of several thousand times—e.g., solving a 10 000‑node instance in 1.48 seconds (≈8600× faster) and a 50 000‑node instance within a few hours—demonstrating that the RNN‑inspired framework can efficiently handle inventory optimization at scales previously unattainable.

The study concludes that the combination of recurrent‑network gradient techniques, tensorized computation, and sparse matrix handling provides a powerful, scalable solution for large‑scale supply‑chain optimization and can be extended to other networked systems such as transportation or service networks.