Causal Bayesian Optimization via Causal Bandits: Open Questions
Arpan Mukherjee, Zirui Yan, Ali Tajer
Proc. Conference on Uncertainty in Artificial Intelligence Workshop on Causality for Decision Making (UAI CDM), 2026.
Abstract
Causal bandits (CBs) provide a principled approach to the sequential design of interventions (experiments) on causal systems to optimize any desired utility of interest. Implicitly, CB algorithms learn the causal model while performing experiments, to the extent that the model's information is needed for optimizing the desired utility. This paper provides a perspective and framework for viewing and analyzing causal Bayesian optimization (CBO) through the lens of CBs. Despite their distinctions, CBO and CB problems share the same principle: causal structure can be used to share information across interventions and avoid treating actions as independent alternatives. Specifically, this paper first compares the two lines of literature on CBs and CBOs with respect to the roles of graph structures, intervention models, and function-class assumptions, distinguishing graph-theoretic sources of complexity from those arising from posterior surrogate-model complexity. Finally, it identifies open problems in CBO that can be addressed using recent advances in CBs.
BibTeX
@inproceedings{mukherjee2026causal,
title={Causal {Bayesian} Optimization via Causal Bandits: Open Questions},
author={Mukherjee, Arpan and Yan, Zirui and Tajer, Ali},
booktitle={Proc. UAI Workshop on Causality for Decision Making},
year={2026},
month={August},
address={Amsterdam, Netherlands}
}