Generative Distributionally Robust Optimization Framework Introduced
A novel framework known as Generative Distributionally Robust Optimization (GDRO) has been introduced in an arXiv paper (2607.24983). This approach tackles the shortcomings of current DRO methods by permitting any sampleable conditional generator as the nominal model and limiting worst-case distributions to a selected family of conditional generators. A significant advancement is the sampler-Sinkhorn pairing: samplers accurately depict conditional laws, while Sinkhorn divergence allows for the comparison of induced distributions without needing likelihood access, relying solely on sample estimates. This enables a straightforward finite-sample approximation and a differentiable primal-dual implementation. The framework seeks to harmonize model compatibility with adversarial structure, addressing the trade-offs seen in earlier techniques.
Key facts
- GDRO accepts any sampleable conditional generator as the nominal model.
- GDRO restricts worst-case laws to a chosen conditional generator family.
- The sampler-Sinkhorn pairing is the key innovation.
- Samplers represent conditional laws exactly.
- Sinkhorn divergence compares distributions without likelihood access.
- Sinkhorn divergence can be estimated from samples alone.
- The population problem admits a direct finite-sample approximation.
- The framework supports differentiable primal-dual implementation.
Entities
Institutions
- arXiv