Decision-Aware Approximation of Belief Functions for Evidential Combinatorial Optimization

What the math gives to ML

The paper replaces representation fidelity with decision fidelity when compressing a distribution over uncertain costs. Its key transferable asset is a one-point regret bound: under a monotonicity condition, the regret of the decision chosen from a compressed representation is bounded by the compression error evaluated only at the true optimum, rather than uniformly over all feasible decisions. This suggests compressing neural uncertainty outputs, ensembles, particles, or evidential cost distributions specifically for the downstream action selected by a planner. A practical first target is a neural graph model that predicts uncertain edge costs and feeds a shortest-path or routing layer, using regret-aware focal-element pruning instead of Jaccard- or distance-based pruning.