Title: Broadening the Scope of Decision-Focused Learning Through Score Function Gradient Estimation

Abstract:

In many practical optimization scenarios, key parameters remain unknown until deployment, often due to inherent stochasticity or incomplete information, such as uncertain demand or travel durations in logistics. The standard approach involves employing machine learning (ML) models to estimate these parameters by minimizing prediction error. However, this objective does not always align with the error metrics relevant to the downstream decision-making process.

The decision-focused learning (DFL) framework addresses this misalignment by optimizing directly for task-specific losses, such as regret. Yet, because combinatorial problems often yield non-informative gradients for these losses, current state-of-the-art DFL techniques rely on surrogates and approximations to facilitate training. These existing methods, however, are constrained by specific structural assumptions, such as convexity or linearity, or by restricting unknown parameters to the objective function alone.

To overcome these limitations, we introduce a novel method that operates without such assumptions. By integrating stochastic smoothing with score function gradient estimation, our approach is applicable to any task loss. This advancement expands the utility of DFL to include nonlinear objectives, uncertain parameters within problem constraints, and even two-stage stochastic optimization. Our experimental results indicate that while the method may require a higher number of training epochs, it matches the performance of specialized techniques. Notably, it demonstrates superior efficacy in challenging scenarios involving uncertainty in constraints, delivering strong results in terms of both solution quality and scalability.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC