Scalable unknown-graph CBO

Develop scalable unknown-graph Causal Bayesian Optimization methods that avoid intractable posterior or confidence-set maintenance over causal graphs as the number of variables increases.

Background

Unknown-graph methods such as CEO and GACBO represent uncertainty over candidate causal structures while optimizing interventions. Their graph representations and updates are demonstrated primarily on small graphs.

As dimensionality grows, posterior or confidence-set maintenance over graphs becomes computationally intractable, motivating local decision-relevant structure learning, amortized graph inference, or implicit graph-aware surrogates.

References

CEO \citep{branchini2023CEO} and GACBO \citep{mukherjee2024} demonstrate unknown-graph CBO on small graphs, but maintaining a posterior or confidence set over graphs becomes intractable as the number of variables grows.

— Causal Bayesian Optimization: Foundations, Methods, and Applications  (2609.24112 - Huang et al., 21 Sep 2026) in Section 6.2, paragraph “Unknown-graph scalability”

However, regret bounds for effect-level CBO with observational priors, for unknown-graph CBO, and for constrained CBO are largely missing. Developing such bounds would clarify when causal structure provably improves over black-box BO and under what conditions.

— Causal Bayesian Optimization: Foundations, Methods, and Applications  (2609.24112 - Huang et al., 21 Sep 2026) in Section 6.5, paragraph “Finite-sample regret bounds”