D-CORL: A provably distributed causal discovery algorithm via ordering-based reinforcement learning

doi:10.1016/j.knosys.2026.116580

D-CORL: A provably distributed causal discovery algorithm via ordering-based reinforcement learning

2026-07-01

0

PRE

AI

DOI：10.1016/j.knosys.2026.116580

Original

Original request for help

Save

Abstract

En

Causal discovery has received significant attention in recent years, which aims to learn the causal structure among random variables in complex systems. However, existing distributed causal discovery methods may suffer from the limitations of large-scale directed graphs searching space or identifying only the Markov equivalence class (MEC) instead of the true directed acyclic graph (DAG). To address these issues, we develop a distributed causal discovery under ordering-based reinforcement learning (RL) paradigm by leveraging Markov boundary and incremental learning techniques, referred to as D-CORL. Specifically, we incorporate distributed paradigm into ordering-based RL framework to explore the ordering, and further infer the optimal DAG structure in parallel to improve efficiency. Meanwhile, the local action corresponded to each agent is selected from a compact candidate node space, which is constructed with only proven dependent nodes of the previous orderings via Markov boundary. Furthermore, theoretical results demonstrate that the D-CORL method can efficiently discover the causal structure in multi-agent systems. Finally, we conduct experiments to verify the effectiveness of the D-CORL algorithm.

Keywords:

Causal discovery

Distributed reinforcement learning

Markov boundary

Directed acyclic graph DAG

Incremental learning

D-CORL: A provably distributed causal discovery algorithm via ordering-based reinforcement learning

Abstract

Journal

Organization

Cited Papers