Beyond the Bounds

In 2004, I chose a computer science faculty position at Brown over the Tepper School of Business for one reason: I was betting that the future of operations research was machine learning (ML). At the time, colleagues dismissed the idea of self-learning optimization as foolish – and those were the ones being polite.

Twenty years later, the same people who attacked the concept are headlining talks on ML for O.R. But they still miss the point. The integration of ML wasn’t just a trick to shave 5% off a runtime. It was a philosophical pivot: solvers should adapt to the real-world problem, not the other way around. ML should help broaden the applicability of O.R., not just make the existing scope an iota more efficient.

For decades, the O.R. community has lamented its lack of impact in the C-suite, blaming poor communication or a lack of data literacy. But the real problem is our obsession with solving the same old problems a fraction faster, rather than augmenting our capabilities to solve entirely new problems. As a result, practically all O.R. vendors rely on the same mixed-integer linear programming (MILP) technology, offering almost no true differentiation. They massage real-world problems until they fit the linear, deterministic world of legacy-solvers. But to have real impact, we need tools built for the messy, non-linear, multi-objective, uncertain, and time-constrained reality 
of modern business.

Bridging the Gap: What Two Decades of Research Taught Us

Getting to a solver that actually “learns” required us to dismantle some assumptions. First, we found that while online learning for a specific problem can improve search, the computational overhead usually offsets the gains. More important, we realized why traditional offline pre-training fails: you cannot devise a representative training set for search decisions because early choices fundamentally change the distribution of all subsequent search states. Optimization must be tuned via end-to-end evaluations.

We also discovered that, while we can learn search parameters for general distributions or specific instances, the real value of ML is not in making a few “big” decisions. Rather, it lies in influencing tens of thousands of low-stakes search decisions. This makes ML a natural partner for fast, iterative primal methods rather than heavy branching logic. We proved this by using hyperparameters augmented by online runtime features to win four out of nine categories in the 2016 MaxSAT evaluation, showing that a self-learned metaheuristic can outperform the world’s best research algorithms.

We also had to address the inherent uncertainty of ML forecasts that feed into optimization models. The common “predict-then-optimize” approach – optimizing for a single point forecast – woefully overfits the chosen scenario and costs businesses millions. Furthermore, the predict-then-optimize trend is a misguided research direction. As very few small examples show, it is theoretically impossible to massage a forecast to bias an optimization model 
such that both risk and expected performance are controlled simultaneously.

From Research to Product: The Birth of InsideOpt Seeker

The only effective way to optimize under uncertainty is to embed the uncertainty evaluation directly into the optimization search itself. Using this approach enabled us to win the 2021 AI4TSP IJCAI competition on prize-collecting TSP under uncertain travel times. That win convinced us to productize this technology to enable non-convex, multi-objective, and stochastic optimization for industry.

The final piece was the modeling front-end. We needed to enable modelers to create simulators of their systems that were tailor-made for optimization. This required an API that allowed the backend hyper-reactive search to re-simulate systems thousands of times per second. 

But the challenge was more than technical. We had to build a company to enable this development. We invested heavily ourselves, and we needed to find additional investors. We had to establish a presence and inform our community about the existence of a viable add-on to mixed-integer programming. We had to learn to do more than build a solver; we had to know how to write end-user license agreements, run modeling workshops, deal with procurement – the list goes on.

Seeker 1.0 went public in October 2023. Now in its sixth generation, our technology allows businesses to balance multiple KPIs while controlling risk, thus solving problems that were previously impossible to tackle while exploiting distributed compute clusters. Notably, the Seeker API allows very fast prototyping, which is essential to support fast turnaround times with the business during the discovery phase. For O.R. experts, this creates a new opportunity to distinguish themselves. Seeker does not replace MILP. Rather, it extends our capabilities in a world that yearns for efficiency. 

Operations research should have more impact than any other department in any organization. If yours doesn’t, the bottleneck is not your people, but your tools. The time has come to differentiate your optimization shop and to offer optimization support where it was lacking before. 

Augment your toolbox with solver technology that actively supports rapid prototyping and enables you to solve linear to non-convex, single to multi-objective, and deterministic to stochastic problems, while effectively exploiting fully distributed compute clusters at scale. The technology to do so exists. It is called AI-based search. Let’s move beyond the bounds together. 

References

Ansótegui, C., Pon, J., Sellmann, M., and Tierney, K., 2017, “Reactive Dialectic Search Portfolios for MaxSAT,” Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 3, No.1. https://doi.org/10.1609/aaai.v31i1.10660.

InsideOpt Tutorials, 2024, “Optimality Is a False Idol [Video],” YouTube, https://www.youtube.com/watch?v=iCqz1bLTDJU.

InsideOpt Tutorials, 2026, “Why Predict-then-Optimize and End-to-End Learning Won’t Fix Your Optimization Under Uncertainty [Video],” YouTube, https://www.youtube.com/watch?v=_Xa085jbYCI.

Kadioglu, S., Malitsky, Y., Sellmann, M., and Tierney, K., 2010, “ISAC—Instance-specific Algorithm Configuration,” Proceedings of the 19th European Conference on Artificial Intelligence, pp.751–756, IOS Press, https://dl.acm.org/doi/10.5555/1860967.1861114.

Leventhal, D. H., Sellmann, M., 2008, “The Accuracy of Search Heuristics: An Empirical Study on Knapsack Problems,” Integration of AI and OR Techniques in Constraint Programming for Combinatorial Optimization Problems, pp.42–157, Springer. https://doi.org/10.1007/978-3-540-68155-7_13.

Malitsky, Y., Sellmann, M., 2009, “Stochastic Offline Programming,” 2009 21^st IEE International Conference on Tools with Artificial Intelligence, pp.784-791, https://doi.org/10.1109/ICTAI.2009.23.

Malitsky, Y., Sellmann, M., 2010, “Stochastic Offline Programming,” International Journal on Artificial Intelligence Tools, Vol. 19, No. 4, pp. 351–371. https://doi.org/10.1142/S0218213010000236

Zhang, Y., Bliek, L., da Costa, P., et. al., 2023, “The First AI4TSP Competition: Learning to Solve Stochastic Routing Problems,” Artificial Intelligence, Vol. 319, article 103918, https://doi.org/10.1016/j.artint.2023.103918.