Decision analytics for supply chain management

Warren B Powell | Chief Analytics Officer, Optimal Dynamics | Professor Emeritus, Princeton University

For years supply chains have grown in complexity, a byproduct of the increasing complexity of what we make and the dramatic reduction in shipping costs, making it economical to reach out to least-cost suppliers from around the world.? Combine this with the increased transparency of prices from e-commerce, and manufacturers are all but forced to find the lowest cost producers in order to get the best prices.

It all worked incredibly well, up until COVID.

The meltdown in supply chains has been swift and unexpected.? The poster child of the problems with supply chains has unquestionably been the piling up of containers into and out of ports, but the reality is that it has affected every dimension of supply chains.? The surge in demand for physical goods has been dramatic, creating shortages in raw materials, component parts such as chips, people, truckload trucking capacity (notably drivers), manufacturing capacity, assembly line workers, final-mile drivers, … did I forget to mention people???

There have been many suggestions for how to fix the supply chain problem, but it is easy to trivialize what is a fundamentally complex problem.? A common suggestion has been to onshore manufacturing to reduce the dependence on China and other distant producers.? This has ignored the vulnerability of even local producers to COVID restrictions and the same shortage of materials and component parts that affect everyone else.

A consistent theme has been how brittle supply chains became during periods of stability.? Manufacturers optimized production to meet the needs of a cutthroat market, focusing on least-cost providers with streamlined inventory and production processes.? Lean production, adapted from Toyota’s pioneering manufacturing style in the 1970s and 1980s has spread throughout manufacturing.? The result is incredible efficiency, but highly vulnerable to disruptions.

Compounding the emphasis on cost has been the emergence of powerful planning technologies in the 1990s under the umbrella of “optimization.”? These tools produce least cost solutions, but they are optimizing simplified and artificial models of real problems.? In particular, they are astonishingly good at finding, well, least cost solutions.??

What is missing from these optimization models is an understanding that the algorithms are optimizing a model of the problem.? A highly simplified model.? The most common approach for modeling the problem of where to locate suppliers is to model average flows from each supplier to meet average demands at least cost (it is easy to limit the search to suppliers that meet various qualification requirements).? What is missing is any representation of uncertainty.??

Before we even get to the challenge of how to make decisions in the presence of uncertainty, we have to identify the different sources of uncertainty.

What types of uncertainty do you have to deal with?

A former secretary of defense, Donald Rumsfeld, preparing for the second Iraq war, talked about “known knowns,” “known unknowns,” and “unknown unknowns.”? The Fukushima disaster, the Suez Canal pileup, and, pre-2020, the COVID pandemic, might all be viewed as “unknown unknowns” when decisions were being made about supply chains that would be impacted by these events.

Our challenge is to identify as many unknowns as possible.? This does not mean that we could have predicted any of the events above.? Instead, we want to identify the sources of uncertainty so we at least can think about events that might happen.??

One company asked me how to do this, and my suggestion was to get a few six-packs of beer and huddle around a whiteboard.? So, grab a beer - we are about to huddle.

In my new book (https://tinyurl.com/RLandSO/), Chapter 10 describes 12 different categories of uncertainty for sequential decision problems (they are listed to the right).? Using these as a guide, I might propose an initial list for supply chain management:

Demand for the product.? Uncertainty comes in several flavors:

• The day-to-day fluctuations of a large market purchasing a product.
• The longer-term variations as market tastes shift, and in response to competitor behavior (pricing, marketing, product availability and quality).

Production capacity of each supplier for each component. This may be due to:

• Supplier shifting capacity to other products/markets.
• Labor problems (including pandemic shutdowns).
• Equipment problems.
• Weather problems.
• Explosions (within or outside of the manufacturing process).
• Manufacturing delays from upstream suppliers.

Costs of components and raw materials.

Currency fluctuations.

Forecasting uncertainties (demand, prices, availability of raw materials).

Theft.

Implementation uncertainties (the field does not implement the plan correctly).

Errors in the reporting of inventory.

Communication errors (true story - a major supplier of diapers nearly shut down their entire production of diapers due to a miscommunication that would have cut off a Chinese supplier before a replacement had come on line).

Shipping delays (sea and land) due to:

• Weather.
• Port delays.
• Equipment problems.
• Piracy (!)

Component/product quality from suppliers.

Availability/price of raw materials and component inputs.

Availability of shipping containers.

Availability of truckload capacity, performance of truckload carriers.

Changes in regulations:

• From foreign governments affecting upstream suppliers or the demand for product.
• Tariffs.

Grab another six-pack and invite someone else to the discussion, and I am sure we can add to this list.

Before we understand how these uncertainties affect our decision-making, we have to apply the same exercise and list the decisions we are making.

What decisions are you making?

The way to run a better supply chain is to make better decisions.? I cannot count the number of times I have asked companies if they have a list of who makes decisions, and what decisions each decision-maker makes.? I just get blank stares.

If you want to run a better supply chain, you need to make better decisions.? You have to start by listing the decisions.??

I claim this is not as hard as listing different sources of uncertainty, but for complex problems such as supply chain management, it is hard.

It is important to remember that decisions come in different flavors.? I like to first divide decisions into three categories:

Decisions affecting physical resources.? Examples are:

• People.
• Physical facilities.
• Fixed equipment (e.g. manufacturing processes).
• Mobile equipment (trucks, tractors, containers, ships, aircraft, ...).
• Contracts for space on transportation capacity.
• Contracts for manufacturing capacity.

Decisions affecting financial resources:

• Working capital on hand.
• Investments.
• Pricing.?
• Insurance contracts.
• Pricing hedges.

Decisions to acquire or share information:

• Investments in sensors (RFID chips, GPS trackers, ….).
• Sales/marketing.
• Laboratory experiments.
• Field experiments.
• Sending/sharing information with vendors/suppliers, the market, and social media.

It is then important to recognize that decisions can be made over a wide range of time periods.? It helps to identify three classes of decisions from this perspective:

• Strategic planning decisions - These are decisions that take place far enough into the future that they do not depend on the state of the system now.
• Tactical planning decisions - These are decisions that impact the near future (this could be tomorrow, a week or two, or months) that do depend on the state of the system now.
• Real-time execution - These are decisions that are executed now.? Typically we want to understand the downstream effect of the decisions.

Closely related to the time frames are the level of an organization where decisions are being made:

• The C-suite - These tend to be strategic decisions
• Middle management - “Middle” covers a wide range of levels in large organizations (I once worked with a shipping company where I could name seven levels of management without naming a vice president).? This tends to cover every decision other than those made by the C-suite, or field execution.
• Field execution - These are the minute-by-minute decisions that are directly implemented in the field, ranging from assigning drivers to loads, firing up machines, and working with customers.

Before we get to the problem of making decisions, we have to first address how we evaluate how well we are doing.

What are your performance metrics?

Understanding performance metrics, especially for larger companies, is like peeling an onion.? At the center is:

• Profits

Since the goal of any private company is to maximize profits, we could ask: what else do we care about?? Profits determine the success of a company, both in the eyes of the marketplace (for publicly traded companies) and employees (at least upper and middle management) whose compensation may be keyed to profits.??

The problem with profits is that for most people, the effect of their decisions is lost in the many elements that determine company profits.? For this reason, companies generally identify a series of high-level metrics that are one level below profits:

• Costs - These of course can be broken down into many categories, aligned with the responsibilities of individual managers.? Note that it is not possible to do the same breakdown with profits (although people try).
• Productivity - Often it is easier to estimate the productivity of people and equipment, rather than cost.
• Revenue - This can be further divided into revenue by product type and product.
• Customer service - Where possible, we can measure how well we are serving customers.
• Employee management - This might be measured in terms of productivity, and (especially important today) employee retention.

Below these levels are dozens, even hundreds, of key performance indices (KPIs) that can measure activities such as waste, cleanliness, employee education, and so on.? In fact, setting these KPIs can represent important decisions made by middle managers to guide decisions in the field.

Finally, it is important to recognize that these metrics can be evaluated over a range of time periods, although the most typical are weeks, months, quarterly and yearly.??

How are you making decisions?

We are now in a position to ask the question: How do we make decisions??

Decision-making is often viewed as a uniquely human activity, with all the inherent complexity of human judgment, logic and thought processes.? Tackling these decisions is typically approached under the banner of “artificial intelligence” (or “AI”), but today, “AI” generally refers to different forms of machine learning, notably image recognition, voice recognition, pattern recognition (all these use neural networks), along with general estimation problems such as demand prediction and forecasting.

We are going to move past this into the next level of analytics: the decision layer.? Machine learning depends on massive datasets to learn patterns (or behaviors) to create functions that take information that we know to estimate values and quantities that we do not know.? There are three classes of functions used in any form of machine learning: lookup tables, parametric models (these may be linear or nonlinear), and nonparametric models. At the heart of all machine learning models is a training dataset.??

Decision analytics does not use a training dataset.? Instead, it uses a physical model of the underlying problem.? To make decisions about deciding which loads to accept, and which drivers to assign to different loads (along with any other decision) we need to design a function called a “policy.”? There are four classes of policies:

1. Policy function approximations (PFAs) - These are analytical functions that include all of the functions used in machine learning (lookup tables, parametric and nonparametric functions).? PFAs are the simplest, such as “if the inventory falls below one level, order up to another level.”
2. Cost function approximations (CFAs) - There are many decisions that are naturally formulated as an optimization problem, such as choosing the vendor with the lowest price, choosing the shortest path, or optimizing the routing of aircraft for major airlines.
3. Policies based on value function approximations (VFAs) - Here we maximize the contribution or reward from an action, plus an approximation of the value of the state of the system that the action takes us to.
4. Direct lookahead approximations (DLAs) - We may choose a decision by optimizing over a planning horizon.? A good example is google maps, where the decision to turn left or right is based on a planned path all the way to the destination.

These four policies can be divided into two broad classes.? PFAs and CFAs (which tend to be the simplest policies) both require tuning parameters, such as the point at which you order more inventory, or (for airlines) the amount of buffer time they allow for weather delays.

VFAs tend to be the most complex, which may explain why they have attracted so much attention from the research community.? Finally, it helps to divide DLAs into two categories.? The first uses a deterministic approximation of future events (as is done by Google maps, which uses a point estimate of the time to traverse each segment of a network), while the second explicitly models uncertainty in the future (producing what is arguably the most complex class of policies).

PFAs, CFAs and deterministic DLAs are easily the most widely used policies in practice.? Policies based on VFAs, and DLAs that explicitly model uncertainty in the future, require more advanced technical skills.

How to choose? This takes some experience.? There is a tendency, of course, to lean toward the policies that are simpler to formulate and solve, but I have consistently found: the price of simplicity is tunable parameters, and tuning is hard!??

An in-depth discussion of the four classes of policies, and a discussion of how to choose among them (including a discussion of hybrids) is given in chapter 11 of my new book.

From order-up-to to direct lookaheads for inventory planning

A nice illustration of the different classes of policies arises when managing inventories.? 60 years of academic research supports the optimality of the familiar “order-up-to” policies: When the inventory falls below one level, order up to another level.? The only challenge is finding the two levels.

I have had top academics almost shout at me: But this is optimal!

Simply put: it is not optimal.? It is only optimal if the demand patterns and costs do not vary over time.? It does not apply with seasonal surges, or even when there is product arriving in large batches on, say, container ships from China.? In fact, this policy ignores the pattern of inbound arrivals.? Remember that lead times from China may extend over 100 days.? How much we should order now depends on what we have ordered previously, and when.? It also matters if we anticipate delays in the port, or due to weather (that we may not have known about when we placed the order, but we do know about when we place later orders).

In addition, we need to consider not only all the different sources of uncertainty (production delays, delays in the originating port, weather delays, currency fluctuations, quality problems, ...) we also need to think about how we might respond to this information as it arrives.? For example, we might be able to air freight product to minimize inventory shortfalls.

Lookahead policies are needed when we have complex, time-varying problems, but you will not find these in any textbooks.? Perhaps it is not surprising that we see these delays in ports since shippers are making decisions using tools based on theory developed a generation ago.

If you want to handle complex problems and different sources of uncertainty, you are going to need methods for making decisions that take these complexities into account.??

Implementing decisions

The transition to computer-assisted decision-making introduces the dimension of using computers instead of people.? However, there is a particular process that eases that transition which involves starting with strategic planning.

• Strategic planning - Strategic planning systems typically involve simulating lower-level decisions to help people evaluate more strategic decisions.? People are still making these decisions, while computers are simulating more detailed decisions that are too tedious for people to estimate on their own.
• Tactical planning - Decisions made now to be implemented in the future (such as ordering inventory from China that arrives in three months, or committing to move a load five days from now) require simulating events between now and when the decision is implemented.??
• Real-time execution decisions - To make the best decision now, it may help to simulate the downstream impact of the decision.

Implementation always opens up a host of issues, notably making sure you have access to the right information. Tactical planning and real-time execution are the most demanding in terms of data because it requires knowing the state of the system: what orders have been implemented, where are inbound orders, what is the state of current inventories (of component parts and finished goods).? These information needs have sparked the creation of visibility platforms.

It is for this reason that it is easiest to start with strategic planning, but this raises the question: how to model complex problems such as supply chains?

Modeling supply chains: deterministic or stochastic?

Most strategic planning models in supply chain management use deterministic approximations that at best capture average flows.? This approach makes the models easy to solve, and still provides input to strategic decisions such as where to choose suppliers, where to store inventory and how to perform shipping (Container shipping or air? Truck or rail?).? It is very common to use deterministic models.

The weakness of deterministic models is that they limit the questions you can ask, and ignore important issues such as the vulnerability of the supply chain to sources of uncertainty.? For example, it should be no surprise that a deterministic model may choose clusters of suppliers with low labor costs in China, ignoring the costs associated with the uncertainty of long lead times.? Sharing production across two suppliers will always cost more than a single supplier.? Rail will always be less expensive than trucking since the variability in transit times is ignored (cross country shipments in the U.S. can arrive a week late).??

Deterministic models not only ignore uncertainty, they also ignore how a company might respond to uncertainty.? Inventory can be staged at different points of the supply chain.? Product can be rushed via air freight if an order is delayed.? Some product problems might be fixed on-site.? Customers might be convinced to accept a substitute product (or a delay in delivery).? This interaction between uncertainty and a variety of contingency actions adds to the richness of supply chain management.? We note that knowledgeable managers are familiar with these strategies.? What is missing is representing them within strategic planning models.??

Closing notes

We have come a long way with deterministic thinking, but recent events are making it clear that we have to start learning to plan into an uncertain future.? Sadly, the root cause of our failure to develop more advanced tools starts with the academic community - after all, we are the ones teaching MBAs and the people with analytical skills.? Today, there is not an academic program in existence that provides students with the tools to solve these complex decision problems in the presence of the uncertainties that modern supply chains have to manage.

A career working on these complex problems has taught me first of all how to think about these problems, which means how to model fully sequential decision problems.? Then, after working on a wide range of applications, I learned that there are four fundamental strategies for making decisions (I call these the “four classes of policies”).? This theory was developed over 40 years at Princeton University, and is summarized in the book to the right (available for pre-order on Amazon right as this is being written - due out in May, 2022). See https://tinyurl.com/RLandSO/ for an overview. These tools are now being put into practice at Optimal Dynamics.