How Advanced Optimization and Simulation Algorithms Transform Supply Chain Planning

Background Overview

Supply chain planning is a strategic upstream activity; adjustments affect downstream processes, requiring global evaluation of cost, efficiency, and experience metrics across many data points and analysis dimensions, which manual analysis cannot handle efficiently.

Intelligent Planning Algorithms and Applications

These algorithms aim to optimize cost, efficiency, and experience by combining operational models, high‑performance optimization, simulation, and intelligent diagnostics, enabling businesses to compute complex solutions, explore optimal combinations, and quantify the impact of strategies.

Warehouse Network Planning

Effective network design improves efficiency and aligns with business models and competitive strategies. JD.com addresses two main scenarios: trunk network site selection and front‑mile warehouse placement.

Trunk network planning optimizes nationwide or regional multi‑level warehouse networks for global cost‑time optimality, balancing handling frequency, distance, and cost to determine product flow, candidate sites, network hierarchy, coverage, and transport modes.

To handle diverse strategies and constraints, a simulation‑driven model generates scenarios (product × node × level × transport) and decouples decision logic via a 0‑1 integer programming model, ensuring generic applicability.

For large‑scale problems with millions of variables, a pre‑filter, variable clustering, and decision‑space reduction algorithm delivers optimal results within 90 seconds, outperforming traditional solvers such as SCIP or Gurobi at this scale.

Front‑mile warehouse placement focuses on single‑city, single‑layer networks serving instant retail and offline stores, achieving full coverage for JD’s internal instant‑retail and offline scenarios.

A GA + Rollout algorithm combines genetic algorithms with reinforcement‑learning‑style exploration, introducing a greedy operator to escape local optima. Compared with open‑source solver SCIP, it improves computation speed by 9‑15×, and shows comparable or superior performance to commercial solver Gurobi, with up to 48× speed gains in some cases.

Inventory Layout

Inventory layout links network planning and replenishment, solving product placement within physical warehouses and transport networks to meet experience, cost, and efficiency goals. The algorithm efficiently schedules millions of SKUs and dynamically adjusts product networks.

A generic selection‑strategy model decouples strategy factors, decision logic, and application, allowing customizable strategy libraries, constraint‑driven selection, and integration with production systems or simulation for rapid, automated responses.

Supply Chain Simulation

Simulation provides quantitative evaluation of business decisions, capturing the “butterfly effect” across upstream and downstream links where A/B testing is costly or infeasible.

The simulation framework builds a sandbox model of the full supply‑chain, combining network structure, product layout, inventory policies, and fulfillment strategies. Optimization and deep‑learning gradient methods search for optimal scenario combinations, while matrix and GPU acceleration boost computation 10‑100×.

Conclusion and Outlook

By leveraging intelligent planning algorithms, JD.com reduces costs and enhances efficiency. Future developments will integrate large language models and reinforcement learning to create more efficient, high‑quality applications, improve algorithm interpretability, lower adoption barriers, and advance the digital intelligence of supply chains.