Generative Distributionally Robust Optimization

What the math gives to ML

The paper provides a concrete way to perform likelihood-free, conditional distributionally robust optimization by restricting adversarial distributions to the output family of a trainable generator and measuring deviation from a nominal generator with Sinkhorn divergence. The transferable asset is the separation between a sampler, which can be arbitrary and implicit, and a sample-based optimal-transport constraint, which supplies a differentiable adversarial objective without densities or scores. A practical neural-network adaptation is to attach a conditional adversarial generator to a predictor or policy, optimize its parameters inside a Sinkhorn-radius ball around the nominal generator, and train the predictor against the resulting worst-case samples.