CDS&E-MSS: Causal Induction in Sequential Decision Processes

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Causal induction, i.e. learning the causal relations among a set of variables from data, is a fundamental problem in scientific research and engineering. An ideal approach to causal induction and inference is through experimental interventions. In many applications, scholars are often interested in maximizing certain outcome variables over a set of candidate experimental interventions, as in the so-called causal bandit problem. Motivated by a few pressing challenges faced by existing methods, namely unknown causality, data heterogeneity, and how to design adaptive experimental interventions, the PI will develop novel statistical methodology and theory for causal bandit and induction based on directed acyclic graphs, a class of popular graphical models for representing causality. Software packages will be released to provide efficient implementation of these methods and algorithms. The research will be integrated into education and research training at both undergraduate and graduate levels. Open-source software with detailed documentation will increase the visibility and the practical application of the research project. This project investigates the causal induction problem under the framework of a sequential decision process. To handle unknown causality, the PI plans to develop a bandit algorithm with causal parent identification and a Bayesian backdoor bandit methodology that averages over parent set uncertainty. Both methods make efficient use of observational data to complement experimental data generated during sequential interventions. This will substantially generalize the causal bandit methodology and theoretical guarantees to the challenging setting under unknown causality. Borrowing the principle of optimism in the face of uncertainty, the PI further develops novel data-adaptive intervention procedures that directly target causal induction, to achieve active learning of the underlying causal graphical model. Since data generated under different experimental interventions do not follow an identical distribution, the PI will develop a new method to learn the interventional equivalence class on such heterogeneous experimental data. The PI will develop theoretical results to establish upper bounds on the cumulative regret in the causal bandit methods and the consistency of the structure learning algorithms. The sequential intervention design for active causal learning is adaptive to finite-size data and can be applied to a general underlying causal graph. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.