Fusing Machine Learning and Mixed Integer Linear Programming
summary
Claim: combining ML and MILP can be a superpower or a disaster depending on details. Examples: the US electricity grid optimization (≈$400 billion/year) and routing and supply‑chain planners; methods discussed include XGBoost, random forests, regression trees, OLS and Bayesian models feeding Mixed Integer Programming. Failure modes include miscalibrated predictors, discontinuous tree outputs, concept drift, and feedback loops that change the data‑generating process. Fixes include calibration or model replacement, monotonicity priors, and using sampled predictions to stress‑test the MIP. Also: marketing mix models and digital twins discussed.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:59Combining ML with optimization
- 4:27Optimization in society
- 6:57Electricity grid challenges
- 9:24How machine learning and optimization complement each other
- 11:34Prediction and optimization workflows
- 13:41ML for faster optimization
- 14:23AI-assisted optimization
- 15:26Digital twins for modeling
- 16:30Price optimization tool
- 17:35ML pitfalls in optimizers
- 20:50ML prediction reliability
- 21:55Regression tree uncertainty
- 24:26Concept drift impact
- 27:16Feedback loops in models
- 29:04Correlated predictors effects
- 32:37Key challenges and solutions
- 33:00Q&A
related talks
17mSynthesizing MILP Constraints for Efficient and Robust Optimization
47mRecent trends in combinatorial optimization augmented machine learning: A graph learning perspective
19mML ain’t your only hammer: adding mathematical optimisation to the data scientist’s toolbox
45mLLMs can't optimize schedules, but AI can!
06mML for Systems and Systems for ML
01h 01mML for discrete optimization: Theoretical foundations
01h 10mMulti-Dimensional and Non-Linear Mechanism Design and Approximation II
35mOn Machine Learning for Prediction and Prioritization in the Allocation of Scarce Societal Resources