**The supply chain works until there is no toilet paper on the shelf****.**

**What is a Supply Chain?**

A simplified supply chain lives in three dimensions: what, where, and when. What is manufactured? Where does it go? When does it get there? This seems relatively simple…until there is no toilet paper on the shelf.

Getting the right items in the right quantity to the right places is complicated. Most organizations manufacture the same product in as few as 5-10 places globally, but want to distribute to every single person on the planet. In the world of mathematics, this distribution challenge can be represented as a series of combinatorial optimization problems. Problems like determining the best routes for shipping and optimally loading shipping containers need to be solved. Few supply chain managers have PhDs in combinatorial optimization or the time to work such difficult problems. Worse, events occur and re-planning has to happen frequently—generally daily. Enter Artificial Intelligence.

**Solving Product Deployment with Artificial Intelligence**

Instead of broadly talking about the applicability of “AI” in the supply chain space, here is an example of how AI can be used in the practice of Product Deployment. Product Deployment is the process typically owned by a “supply chain planner” in which they determine what goes where and when. For example, which distribution center gets toilet paper when it comes off the line? Artificial intelligence can facilitate this, but there are three challenges to address:

**1: Forecasting**

To optimally solve a given combinatorial system like the “what, where, when” paradigm, there first needs to be an understanding of the current and future reality. To do this, forecasts are required that accurately represent demand in each region (this is often represented by a single distribution center in a supply chain). These forecasts take into account existing orders, expected orders, historical trends, and seasonal trends. For each region, a separate forecast will be created, detailing what a supply chain should expect over the next week, month, and quarter. In an ideal scenario, these forecasts also produce confidence levels, indicating a range of possibilities as time goes on.

One limitation of these forecasts is that they will not predict “Black Swan” events like COVID-19, causing a “run” on toilet paper, or other unknown unknowns. Forecasting technology is designed to learn from history and trends to accurately predict demand in steady state and seasonal operations. This steady state forecasting process can be automated and presented to a supply chain planner for final analysis.

Most ERP tools, such as SAP and Kinaxis, have built-in forecasting capabilities, but other technologies, from vendors like SparkCognition and Amazon, have specialized toolsets that leverage deep learning. Deep learning forecasting methods learn intricate relationships in the data and frequently predict future states much more accurately than traditional (statistical) forecasting methods.

**2: Deployment Optimization**

Once there is an accurate depiction of how many goods are going to be purchased in the future, organizations can start to decide how to best act upon this information. When actually deploying goods, there are a variety of constraints that need to be considered across the supply chain network. Some examples of these constraints are:

- Goods must arrive at distribution centers prior to purchase
- Each manufacturing facility has a limited production, storage and shipping capacity
- Trucks, intermodal and railcars have different travel times depending upon final destination
- The price of transportation can fluctuate based on demand, contracts, the spot market, and oil prices
- Each distribution center has a capacity of goods it can receive and the number of trucks it can unload

When all the different products and constraints are considered there is a massive optimization problem, like a sudoku puzzle with 1,000,000 squares. Most supply chain planners don’t have the time or capability to solve this puzzle on a daily basis, so they find an approximate solution that at least tries to get the product to the target destination BEFORE it is needed.

By leveraging optimization technologies like linear programming, better solutions can be found that account for all of the different constraints within the network. Linear programming works by taking complex problems like “what, where when?” and breaking them down into simple linear equations that can be solved simultaneously to achieve the best global solution. Unfortunately, these technologies are also imperfect. They find an approximate solution quickly using solvers like IBM CPLEX or Gurobi Optimizer, but their solution often requires fine tuning. For example, optimizers can produce sound but impractical results like partially loaded vehicles on a lane.

**3: Container Loading**

To resolve this optimizer shortcoming, the final technology needed is full container loading. Combinations of products with varying weights and dimensions are often not loaded perfectly (or practically) by the Deployment Optimizer. There needs to be a technology that can decide how to best load prioritized pallets of goods into a variety of different containers. In the combinatorial mathematics world, this is often referred to as the knapsack (weight-based) or bin packing (volume-based) problem.

Unfortunately, the real world isn’t as straightforward as combinatorial mathematics academia especially when each mode of transportation has its own complexities. When loading trucks, considerations must be made about volume, axle weight distribution, and stack-ability. Ultimately, the supply chain planner will need to turn to a complex, customized, solver to determine how to load goods into a set number of containers for each lane. This solution needs to be fast, and is often integrated as a sub-component of the Deployment Optimization process. Bleeding edge technologies for container loading like reinforcement learning are making this integration possible, with groups like ProvisionAI and Alibaba beginning to embed this capability into their solutions.

**The Future of AI in Supply Chain**

The life of a supply chain planner will improve as artificial intelligence helps to optimize their complete network daily. However, Product Deployment is only the tip of the iceberg. There are other practical AI applications like warehouse orchestration and raw material sourcing that have yet to be fully realized. Until all segments of a supply chain are viewed as synchronous processes that fit together, there will always be room for AI.

##### Keith Moore

Keith Moore is a co-founder of ProvisionAI and the Chief Product Officer for Transportation | Warehouse Optimization - AutoScheduler. He spends most of his time working with his talented teams on developing machine-learning software for better orchestrating components of the supply chain. Moore was voted by Hart Energy Magazine as an Energy Innovator of the Year in 2020.