Towards MFACBO: Multi-Fidelity Abstraction Causal Bayesian Optimization in the Context of the Abstraction-Fidelity Connection

Jakob Zeitler

Keywords: causal abstraction, causal bayesian optimisation, multi-fidelity

Abstract: We investigate a novel integration of Causal Bayesian Optimization (CBO, \cite{aglietti2020causal}) and Multi-Fidelity Bayesian Optimization (MFBO, \cite{matthias2017}) at the intersection of Causal Abstraction. MFBO enables cost-effective exploration by pooling information from fidelities with differing costs, while CBO introduces structural assumptions through incorporating causal knowledge—particularly Directed Acyclic Graphs (DAGs) encoding intervention relationships that can enhance multi-fidelity optimization. This fusion, which we term Multi-Fidelity Abstraction Causal Bayesian Optimization (MFACBO), is expected to improve decision-making efficiency in resource-constrained settings, such as healthcare or physical simulation, by guiding both the selection of intervention sets and fidelity levels. At the core, we expect Causal Abstraction to characterise the relationship between CBO and MFBO, where different fidelities are assumed to exist on differing levels of abstraction. We will evaluate our approach through synthetic experiments and real-world inspired scenarios using the recently introduced Causal Chambers, with particular attention to fidelity correlation modelling and acquisition strategies. This work lays the foundation for a principled and practically motivated framework that integrates causal reasoning into fidelity-aware Optimization, underpinned by the fundamental theory of causal abstraction.

Submission Number: 6