Megadose The AI curve, compressed.

Pareto-Informed Smart "Predict, then Optimize" for Multi-Objective Combinatorial Problems

· ArXiv · AI/CL/LG ·
Training against Pareto-frontier decisions beat both prediction-error training and scalarization-based SPO+ in the authors’ tests.

Aabhash Dhakal and Tim Lachner extend decision-focused learning to combinatorial problems with conflicting objectives. Their loss uses the true Pareto frontier to select a target decision aligned with a sampled preference vector, then minimizes regret against that target. Experiments cover multi-objective shortest-path and knapsack problems. The abstract reports better performance than both baselines but gives no numerical margins.

ArXiv · AI/CL/LG's note

score 4

Categories: Research