ICML 2026 Breakthrough: Solver-Free Method Cuts Predict-Then-Optimize Training by Orders of Magnitude
A new training method eliminates solver calls, making decision-focused learning scalable and fast.
In predict-then-optimize, machine learning models predict coefficients for downstream linear or combinatorial optimization tasks. A key challenge is training to minimize decision regret directly, because the decision mapping is piecewise constant with zero gradients almost everywhere. Existing methods smooth the differentiation process but require computationally expensive solver calls for each gradient evaluation, limiting scalability.
The new method from Wan and Liu proposes a measure transformation principle that rewrites the expected decision regret as an integral, leading to a surrogate loss that is completely solver-free during training. This eliminates the need to call an optimizer in the gradient loop, dramatically accelerating training. The authors prove Fisher consistency and excess risk bounds, and empirically demonstrate decision quality competitive with state-of-the-art methods while training orders of magnitude faster.
- Eliminates computationally expensive solver calls during each gradient evaluation, enabling faster training.
- Introduces a measure transformation principle to derive a solver-free surrogate loss for decision-focused learning.
- Achieves orders of magnitude reduction in training time while maintaining competitive decision quality on optimization tasks.
Why It Matters
Unlocks scalable decision-focused learning for large-scale optimization in logistics, finance, and operations research.