Sampler-Robust Optimization for Generative Models

The recently introduced Sampler-Robust Optimization (SRO) framework tackles reliability challenges in stochastic pipelines that utilize learned generative models. Instead of depending on a static nominal sampler, SRO focuses on optimizing decisions against the worst-case scenario produced by perturbations to the generator. This method is compatible with simulation-based decision-making processes and presents a sharpness-aware perspective, prioritizing decisions that remain stable amid generator variations. Given a coverage assumption, the empirical worst-case objective yields a high-probability upper certificate. The approach aims to address two significant issues: errors from sampler misspecification and those arising from finite simulations.