Intelligent Supply Chain Planning Algorithms and Their Applications

Background Introduction

Supply chain planning belongs to the strategic layer and is an upstream part of the supply chain; adjustments at this level cause massive changes downstream, requiring a global view to evaluate cost, efficiency, and experience metrics across many data dimensions, which would consume huge manual effort.

Supply Chain Planning Algorithms

These algorithms empower the business both in breadth and depth: using operations models and efficient optimization algorithms to compute complex problems and search for optimal combinations, while simulation and intelligent diagnosis help quantify the impact of different strategies and uncover optimization opportunities.

Intelligent Planning Algorithm System

The intelligent planning algorithm aims to optimize supply‑chain cost, efficiency, and experience by building a technology stack that links intelligent diagnosis, digital solutions, execution, and operational monitoring. The following sections introduce the algorithms and application scenarios for warehouse‑network planning, inventory layout, and supply‑chain simulation.

Warehouse‑Network Planning

Reasonable supply‑chain network design improves efficiency and serves the business model and competitive strategy. JD’s internal and external customers mainly need two scenarios: trunk‑network site selection and front‑warehouse network selection.

Trunk‑network site selection is a multi‑level, national or regional planning task that seeks global optimality of cost and timeliness, designing the full‑link network from supplier to customer while considering handling frequency, distance, cost, flow direction, candidate nodes, hierarchy, coverage, and transport modes.

Because many customized strategies lead to low delivery efficiency and high cost, a simulation‑based trunk‑network model was designed. Business data, policies, and constraints generate "product × node × level × transport‑mode" scenarios, and a binary‑scenario‑selection integer programming model decouples decision logic from business logic, ensuring model generality. The solution is comparable to products such as Llmasoft and BlueXing, and is implemented in the NetSim system supporting various network projects across retail, logistics, and industry.

To accelerate solving large‑scale trunk‑network problems, a pre‑filter + variable‑clustering + decision‑space‑reduction algorithm was created. It can handle 80 million decision variables and output the optimal result within 90 seconds, a scale where solvers like SCIP or Gurobi fail to model the problem.

Front‑Warehouse Network Selection

Front‑warehouse network selection is a single‑layer, city‑level planning task serving instant‑retail, offline retail, and service scenarios. It supports a wide range of site‑selection cases (e.g., fresh‑food warehouses, health‑care, quick‑delivery, 7‑fresh, medical, 3C stores, maternity, comprehensive, JD‑car‑service), achieving 100 % coverage of instant‑retail and offline‑retail scenarios.

GA + Rollout Solving Algorithm

To break the computational bottleneck of large online cases, a GA + Rollout algorithm was co‑designed, combining the exploration mechanism of reinforcement learning with classic Genetic Algorithm operators and a direction‑guided greedy exploration operator. Compared with the open‑source solver SCIP, the new algorithm improves efficiency by 9‑15 × while maintaining solution quality; against the commercial solver Gurobi, it shows complementary strengths and can achieve up to 48 × speed‑up in some cases.

Inventory Layout

Inventory layout sits in the planning layer, linking warehouse‑network planning and replenishment, providing solutions for product placement based on physical warehouse and transport network relationships while meeting cost, efficiency, and experience goals. The algorithm covers scenarios such as BBCC, large‑item layout, fashion, books, quick‑delivery, and cross‑border, achieving a supply‑chain cost reduction of over 5 billion RMB.

A generic selection‑strategy configuration model decouples strategy factors, decision model, and application modules. By building a standard factor library, customized strategies can be assembled; the decision module outputs selection results based on user goals and constraints; and the application module integrates with production systems or runs simulation evaluations. This model supports millions of SKUs and enables rapid, automated responses to business demands, with 24 % of inventory‑layout scenarios launched online and three scenarios fully automated.

Supply‑Chain Simulation

Supply‑chain simulation is a quantitative decision‑support tool that captures the tight coupling between upstream and downstream links, where adjustments upstream cause a "butterfly effect" downstream. Physical A/B experiments are costly and slow, and some scenarios cannot be tested; simulation provides a digital sandbox to evaluate strategy impact.

The simulation technology includes model construction, solution search, and computational acceleration. A universal full‑link simulation model combines network structure, product layout, inventory policy, and fulfillment strategy. Gradient‑descent methods from operations research and deep learning optimize solution combinations and policy parameters, while matrix calculations and GPU stacking accelerate simulation by 10‑100 ×.

In the BCC project, the simulation supports cost and rate evaluation for different inbound‑warehouse schemes, covering seven business units, generating over 80 online rate‑simulation solutions, and assisting procurement negotiations. Additional simulations for FDC inventory width/depth, resource planning for major promotions, and VMI timeliness have improved resource assessment and service experience.

Summary and Outlook

The author presented intelligent planning algorithms and how they empower supply‑chain operations to achieve cost reduction and efficiency gains. With the next wave of AI technologies, future work will integrate large models and reinforcement learning, improve algorithm explainability, lower adoption barriers, and jointly advance the digital intelligence of supply chains.

Scan the QR code to join the technical discussion group.