Top 3 AI use cases for supply chain optimization

Top 3 AI use cases for supply chain optimization

“Supply chain” is a simple moniker that encompasses a broad swath of business processes, starting from receiving raw material from the supplier to delivering the finished product to the end-customer. It is a complex network of suppliers, warehouses, manufacturing plants, logistic operators, global national or regional distributors, and retailers. Therefore, companies continuously strive to optimize their supply chains to reduce costs and improve operational agility.

Supply chain production flow and optimization diagramAs the pandemic showed, simple “business as usual,” mostly linear or rule-based optimization algorithms, don’t have the flexibility and reaction time to adapt to quickly changing conditions. This pandemic-induced systemic shock showed that a new approach to supply chain optimization. Specifically, AI-based solutions are increasingly becoming the go-to approach to both increase supply chain robustness and reduce risks and costs. With AI, manufacturers can reduce cost through reduced operational redundancies. They can also mitigate supply chain risk and enhance forecasting accuracy. Finally, they can provide faster deliveries through optimized routes,

However, bringing theoretical (i.e., research-based concepts) AI to real-life business challenges is nothing but simple. Therefore, in this article we will focus on the three most common supply chain bottlenecks and the benefits of applying one innovative and specific approach to AI: Deep Reinforcement Learning (DRL).

three circles with icons for inventory optimization, production planning, and warehouse storage


Why Deep Reinforcement Learning makes sense for supply chain optimization

Taxonomy note: Autonomous Systems are systems that leverage AI agents (i.e., “brains”) that were trained using Deep Reinforcement Learning using, for instance, the Microsoft Project Bonsai platform. In the remainder of this article AI agent, Bonsai brain, Autonomous Systems, and Deep Reinforcement Learning will be used interchangeably based on context.

AI is, theoretically, an appealing technology to manufacturers, distributors, transportation companies, and other player within the supply chain However, one of the perennial issues with AI is the lack of available training data for real-life use cases. AI training data is usually a labeled data set used to train AI models or ML algorithms. Such a large amount of reliable data, from hundreds of thousands to millions of individual data points, is usually not readily available, however.

Deep Reinforcement Learning is a technique that leverages a process simulator to let the AI agent train on its own based on a trial-and-error approach. The agent performs hundreds of thousands to millions of iterations to learn in a sandboxed environment.

DRL is only applicable if it is possible to build an accurate simulator, which is always the first step in such a project. Therefore, it is critical to choose the right simulator adapted to the particular use case. Once the simulator is available, one can fairly easily and quickly use Deep Reinforcement Learning to train AI agents. Those agents are called “brains” in the Microsoft Project Bonsai development platform.

In Bonsai, the platform let the AI agent t train itself for a broader set of (simulated) situations. The agent (aka “brain”) will then be able to adapt to changing conditions. At least if those conditions were within the scope of the simulator when the AI agent was trained. Most importantly, the AI agent does not rely on human operators or data scientists to plan, ahead of time, all the possible inputs or environmental conditions variations. The ensuing AI agent will then be robust to changing parameters (inputs, datasets, and environmental changes) and will easily adapt to the evolving conditions.

To learn more about Deep Reinforcement Learning, you can check out this blog post and this video series.

Diagram of deep reinforcement learning training loop

Top 3 use cases for AI-powered supply chain optimization

Use case 1: Inventory optimization

Many-to-many relationships between distribution centers and local supply locations force complex allocation decisions. Poor regional replenishment decisions can lead to local stock-out conditions and lost sales.

Often, classical methods like linear optimization (aka linear programming) are used. Those are often driven directly off the company’s ERP, such as SAP APO.

However, these simple linear equations struggle to adequately represent real life variability, especially during periods of rapidly changing local demands.

Using Project Bonsai, one can build a brain (AI agent) to dynamically optimize regional inventory replenishment to create more stable and predictable stocking levels, thereby avoiding lost sales.

operator with hardhat and tablet


Use case 2: Production planning

It is difficult to plan production levels with everchanging forecasts, raw material costs, labor constraints, and shipping costs. Often, product change on the manufacturing line is time-consuming and costly if not properly optimized to meet customer demand and inventory needs.

Unfortunately, current methods used by (human) production planners involve current forecast and available market data planning, while manufacturing teams schedule production based on similar, but not always aligned, forecast and equipment availability.

These methods are difficult to develop, but they’re also error-prone when there is a need to simultaneously manage the changing optimization goals and the current production goals.

To improve production planning and solve these limitations, one can build an AI agent using DRL to optimize production by determining amounts of which product SKUs to manufacture and how to best schedule their production.

Boxes in warehouse

Use case 3: Warehouse storage and retrieval optimization

Warehouses store a wide range of products that require different storage and handling strategies. Inefficient storage and retrieval decisions can have severe negative financial impacts.

Current methods often leverage human operators and relatively simple rule-based algorithms to dictate storage location upon loading and retrieval policies upon shipment.

However, these manual methods are static and do not adapt to changing customer and market demands.

Manufacturers can improve both storage and retrieval operations by building an AI agent that can dynamically optimize and balance throughput and efficiency within the warehouse to maximize financial return.

Learn more: