Algorithms: Dimitris Bertsimas Killian Lecture on life and

Dimitris Bertsimas delivered MIT's 2026 Killian Lecture on March 23, 2026, presenting a retrospective of his work in optimization and a forward-looking view of how learning systems can be integrated with decision-making. The lecture covered theoretical advances, software and classroom practice, and examples of applying algorithms to operational problems in healthcare, finance and industry.

Lecture highlights

Bertsimas opened with a survey of optimization theory developed over his career, emphasizing the interaction between tractable models and real-world constraints. He reviewed how mathematical programming and robust optimization provided tools for structured decision problems, then described efforts to marry those tools with statistical learning. He sketched ways models that were once purely prescriptive have incorporated prediction components, and how that hybridization influenced project selection, software design and student research topics.

The talk included concrete classroom and tooling priorities: reproducibility in computational experiments, teaching students to balance model complexity with interpretability, and building software that practitioners can deploy. Bertsimas used several case studies to illustrate these priorities, each linking a theoretical idea to decisions made in operational settings. He stressed the importance of clear decision rules when systems affect patients, customers or supply chains.

Audience questions ranged from technical details about algorithmic scalability to broader themes about career choices for graduate students. The lecture format allowed time for demonstrations of recent software used by his group and for discussion of open problems that arise at the intersection of optimization and data-driven methods.

Research directions and takeaways

Bertsimas outlined several research directions he considers ripe for development. First, tighter integration of learning and optimization, where predictive models are trained with downstream decisions in mind rather than as standalone modules. Second, a focus on interpretable and verifiable decision models that let domain experts understand and validate recommendations. Third, improving the software and computational infrastructure needed to bring research prototypes into sustained operational use.

He also highlighted human-centered concerns: the need for clearer communication between modelers and domain experts, methods for robust deployment under shifting data distributions, and training pipelines that prepare students for interdisciplinary collaboration. Across these themes he emphasized methodological rigor paired with attention to deployment constraints.

Why it matters

The lecture underscores a shift in operations research and related fields toward tightly coupling prediction and decision-making, not treating machine learning as an isolated component. For practitioners and educators, the emphasis on interpretability, reproducibility and deployment signals where funding, coursework and hiring may concentrate in the coming years.