Bauer group
Bayesian Causal Discovery with Optimal Experimental Design
Many scientific disciplines involve perturbing the system of interest to discover and understand the mechanisms that influence the data generating process. A proper mechanistic understanding of any system is enabled through the framework of Structural Causal Models (SCM), with a corresponding causal graph that indicates causal relationships between the variables. However, in many applications, the SCM and the corresponding causal graph is unknown and has to be recovered from data. This problem, called Causal Discovery, is theoretically impossible just from observational data alone. This necessities the need for interventions to obtain interventional data. Given the large space of possible interventions and the cost involved, these interventions need to be planned such that the causal graph can be discovered with as few interventions as possible. However, the design space of possible interventions is discrete and usually large due to combinatorial possibilities of intervention targets. In this talk, I will outline some recent work which addresses this problem by treating the optimal experimental/ intervention design problem as an optimisation problem in a Bayesian setting. This enables optimal selection of multi-target nodes/ variables to intervene on through relaxed gradient estimators.