The Decision Factory

You created the perfect plan the night before...
Reality shattered it by morning.

Sound familiar?

Fulcrum Logistics had a beautifully formulated MILP, brilliant data scientists, and the best intentions. Yet every morning began with a new crisis. Their “optimal” plans couldn’t survive contact with reality: demand spikes, traffic jams, and a thousand small uncertainties that quietly destroyed deterministic solutions.

This is the story of how they fixed it. Not by building a better model, but by building a better decision system.

Through the lens of a team learning to rethink decision-making, you'll discover how to:

• Transform brittle optimization models into adaptive policies

• Build simulators that become laboratories for decision-making

• Use Monte Carlo methods to compute what matters under uncertainty

• Surface tradeoffs instead of hiding them in weights

• Deploy systems that actually work when things go wrong

The Decision Factory teaches sequential decision analytics the way it should be learned: through a story where the failures matter as much as the successes, and the lessons stick because you watched someone earn them.

📖 Foreword From Warren Powell

Very early in my career I was introduced to the problem of managing the fleet for a major truckload carrier by a director with a master’s in operations research.  He described the models that were pioneered at the company and closed with: “But they are deterministic, and our problem is stochastic.”  He then went on to describe the highly dynamic process by which shippers offer loads to carriers over time.

I immediately realized it wasn’t just that I didn’t know how to solve the problem… I didn’t know how to think about the problem.  The dinner conversation launched a career of floundering around the academic literature addressing the problems of making decisions under uncertainty, full of elegant math with little attention to modeling and computation.   

It took me my entire career to get my arms around this rich problem class.  My lab at Princeton, CASTLE Lab, worked on a wide range of problems spanning freight transportation, logistics, energy systems, health, e-commerce, and the countless problems contained in the research papers of 60 graduate students and post-docs, and over 200 undergraduate senior theses.

By 2011 I had published the 2nd edition of my book Approximate Dynamic Programming, and I was starting to realize that ADP simply wasn’t the best approach for making decisions for many (indeed most) problems.  I just had to look at my own projects to realize that different problems called for different methods, known as “policies.”  However, I realized that all of the methods could be organized into four classes (called “meta classes”):

• Policy function approximations (or PFAs) – These are analytical functions, which might be simple rules, linear or nonlinear functions, or deep neural networks.  PFAs are the only class that do not involve solving an imbedded optimization problem.

• Cost function approximations (or CFAs) – CFAs are deterministic optimization models that do not plan into the future which are parameterized to work well over time. CFAs have been completely overlooked in the academic literature, but are widely used in practice, albeit in an ad hoc way.

• Value function approximations (or VFAs) – These are policies that plan into the future by creating an approximation of the value from a decision moving you into a downstream state.  This is where my ADP book fell.

• Direct lookahead approximations (or DLAs) – These are methods that make decisions now by explicitly planning into the future.  We can use a deterministic approximation of the future (as is done in Google Maps), or explicitly model uncertainty, as is done in decision trees.

My major claim is that these four classes include any method for making decisions, including anything in the research literature and whatever is being used in practice.  

Making the transition from applying one of the classes of policies to searching over all four classes represents a fundamental shift in how we think about making decisions over time. This is where The Decision Factory: A Novel About Decisions Under Uncertainty comes in.  Adam DeJans Jr. and John Brandon Elam do an absolutely brilliant job of communicating the idea of using any method for making decisions, whether it is a simple rule or a large integer program.

The style of communication is best illustrated using one of my favorite passages from the book:

``These are just... intuitions,'' she said. ``Things that work.''

``Maya, these are exactly what we've been talking about,'' Alex replied.
``State-dependent decision rules. We've been testing greedy policies and MILP policies in the simulator,
but you've had a whole library of operational heuristics this entire time. Why didn't you say something?''

``I didn't think they counted,'' Maya said. ``They're not... math.''

``They're absolutely math. They're functions that map state to decision.
And unlike our greedy baseline, yours actually encode fifteen years of operational knowledge.''

The book uses this style throughout, highlighting the gap between traditional ways of solving problems (“intuitions”) and advanced tools such as mixed-integer linear programs (MILPs).  

After communicating the concept of “policies,” DeJans and Elam then tackle the challenge of evaluating policies.  The math programming community thinks of making decisions using an objective function to evaluate the decision, where sophisticated algorithms might guarantee that the solution is “optimal.”  However, this is always at a point in time.  Evaluating policies requires looking at how well they perform over time.  Ideally this is done in a simulator, but ultimately this describes how they are evaluated in the field.  

In addition to evaluating policies based on how well they perform on average, they even address the subtle but important issue of understanding worst-case outcomes, which are typically identified as risk.  This is completely ignored not just by deterministic optimization models, but also by the vast majority of stochastic optimization models.  Yet, they still manage to do it with the mechanism of easy-to-understand conversations between managers.

The book does a really nice job of illustrating how simulators can be used both to compare policies, as well as to tune the parameters that almost always exist.  In fact, they even illustrate that a simulator can be a form of policy which simulates the process of making decisions in the future (using what I call the “policy-within-a-policy”).

The approach of searching for the best policies that work well over time means replacing deterministic “objective functions” used by traditional optimization solvers with simulators that embrace uncertainty.  Again, DeJans and Elam make the recognition of uncertainty a natural part of the modeling process, without the mind-numbing complexity of “stochastic optimization.”  

This book has the potential to do for automated decision-making in business that The Goal (Goldratt) did for supply chain management.  The ease with which DeJans and Elam communicate these fundamentally new ideas for making decisions under uncertainty for complex business problems is likely to do more to democratize “stochastic optimization” than anything that has ever been written.  

Warren B. Powell
Professor Emeritus, Princeton University

📖 Why trust the Bit Bros? 🤔

Modelers. Builders. No fluff.

At BitBros, we don’t write academic theory and call it a guide. We write books we wish we had on our desks while building real optimization systems under pressure: deadlines, ambiguity, and all

Our team has helped model supply chains for Fortune 50 companies, taught optimization to hundreds of professionals, and debugged more bad big‑M constraints than we’d like to admit. We know what breaks models, what makes them scale, and where theory usually falls short in practice.