Meinolf Sellmann: Decision-Making Under Uncertainty

Meinolf Sellmann is a computer scientist, entrepreneur, and founder and CEO of InsideOpt. He previously built optimization solvers at Bell Labs, IBM, and GE, giving him a long view of how operations research has evolved from tightly defined mathematical models toward systems that must work under uncertainty, competing objectives, and real operational constraints.

His perspective starts with a conversation he had with George Nemhauser in 1998. Nemhauser identified three major problems for operations research: producing useful decisions within limited time, handling multiple objectives simultaneously, and making decisions when the future is uncertain. Nearly three decades later, Sellmann argues that the same problems remain unresolved. They also remain important because companies and public institutions rarely make decisions based on a single, clear objective with complete information.

Machine learning appeared to offer part of the answer by giving businesses a more systematic way to predict demand, prices, traffic, and other uncertainties. But it could not remove uncertainty. A supermarket may estimate tomorrow’s sushi demand accurately on average and still be surprised by an event that could not have appeared in its historical data. The challenge is therefore not to create a perfect forecast but to make a good decision while accepting that every forecast will leave something unexplained.

Sellmann believes that requires a different type of optimization. Traditional mixed-integer programming solvers rely heavily on dual bounds, eliminating parts of the search space that cannot contain an optimal answer. That approach works well for structured deterministic problems, but it becomes less useful when objectives are nonlinear, uncertainty is material, or several competing outcomes must be balanced. A primal solver starts with possible solutions and repeatedly modifies them, using machine learning to guide the search toward more promising alternatives.

This changes how computing power can be used. Instead of coordinating a small number of tightly linked search processes around a shared bound, a primal system can run many machine-learning-guided search operators in parallel. Sellmann says InsideOpt’s Seeker solver can scale across 1,000 cores and, on the Taillard quadratic-assignment instances, find a stronger solution roughly 1,000 times faster than Gurobi under the benchmark conditions he describes. 

The larger problem is often not computational performance but problem definition. Business customers do not arrive with clean benchmark data, stable constraints, and a single agreed objective. The technical model is always an approximation of the real system, and teams often distort the business problem to fit the solver they already know. Sellmann sees the real skill as deciding where simplification is harmless and where it changes the decision enough to make the result unusable.

A coffee-roasting scheduling project shows the difference. A company used mixed-integer programming to increase utilization across its roasters and cooling silos. The schedule looked strong mathematically, but operators repeatedly deviated from it because adding freshly roasted beans to a silo reheated coffee that had already cooled. That delayed packaging kept silos occupied and disrupted the rest of the plan. The model had optimized a simplified production process rather than the process employees actually had to run.

A primal approach enabled the team to represent silo behavior using a custom function and simulate how coffee would move through the plant. The solver could then search among feasible schedules. 

Sellmann’s own rules for decision-making follow the same logic. Decide only what must be decided now, because additional information may change later choices. Distinguish between reversible decisions and commitments that are expensive to undo. For decisions that create lock-in, expand the available set of actions by adding safeguards, insurance, hedges, or other options that reduce downside risk if the world develops differently than expected.

For Sellmann, good optimization is therefore not about maximizing one expected result. It is about finding a workable compromise across performance, uncertainty, reversibility, and risk. Better decision systems will not eliminate imperfect forecasts or messy business conditions. They will help people act responsibly inside them by representing reality more faithfully, exploring more possible actions, and protecting against outcomes the model cannot predict.

📅 CDAO Fall 2026 (Boston, MA - October 26-27, 2026) 

Now in its 13th edition, CDAO Fall brings together chief data officers, analytics leaders, AI executives, and transformation leaders from organizations across healthcare, financial services, manufacturing, and technology. The program focuses on enterprise data strategy, analytics governance, AI operationalization, and the organizational changes required to translate technology investments into measurable business outcomes. Details →

📅 Gurobi Decision Intelligence Summit 2026 (Las Vegas, NV - September 22-23, 2026)

The Gurobi Decision Intelligence Summit brings together decision scientists, analytics leaders, optimization practitioners, and business executives to explore how organizations are improving decisions through optimization, AI, and decision intelligence. The agenda focuses on customer use cases, technical training, expert panels, networking, and one-on-one sessions where attendees can discuss specific models, constraints, and performance challenges with Gurobi specialists. Sessions will also examine how generative AI can make optimization tools easier to use and help teams turn predictive insights into practical operating decisions. Details →

📊 Report Spotlight: Global AI Pulse Q1 2026 (KPMG) 

KPMG’s Global AI Pulse Q1 2026 draws on responses from 2,110 C-suite and business leaders across 20 countries to examine why rising AI investment is still producing uneven enterprise value. The report finds that the next challenge is connecting AI systems and agents across workflows, functions, and operating models. Organizations making progress are combining technical deployment with stronger governance, trusted data, workforce preparation, and clear human accountability. As agents move deeper into daily operations, the ability to coordinate them securely, redesign work around them, and build employee skills will determine whether AI produces repeatable business outcomes at scale. Read →

Benjamin Baer, Founding Member of the DecideWise community, traces how vendors have long used anatomical metaphors to explain decision intelligence, and what those metaphors clarify and obscure. The oldest is the enterprise nervous system, the idea Tibco built around the enterprise service bus, a middleware layer that connected applications, captured business logic, and pushed real-time alerts and responses. Later came the enterprise brain, seen in Banco Bradesco's reengineering of its systems into a business brain that senses and responds across customer touchpoints, and the notion of decision memory, which Axiom frames as a repository of decision knowledge and history. Baer extends the body analogy further, mapping streaming analytics to a reflex, predictive analytics to judging a trajectory, AI to spotting patterns, simulation to testing scenarios, optimization to choosing the best action, and the learning loop to improving after each outcome. Read More →

Borja Menéndez and Zohar Strinka discuss what analytics professionals need beyond technical modeling skills. Using the Advent of OR challenge and the INFORMS Analytics Framework, they explain how effective decision work depends on problem framing, stakeholder input, systems engineering, curiosity, and repeated model refinement. The episode shows how operations research moves from an elegant mathematical model to a working system that people can actually use.