Reinforcement Causal Structure Learning on Order Graph

Learning directed acyclic graph (DAG) that describes the causality of observed data is a very challenging but important task. Due to limited quantity and quality data, non-identifiability causal graph, it almost impossible infer single precise DAG. Some methods approximate posterior distribution DAGs explore DAG space via Markov chain Monte Carlo (MCMC), over nature super-exponential growth, accurately characterizing whole intractable. In this paper, we propose Reinforcement Causal Structure on Order Graph (RCL-OG) uses order instead MCMC model different topological orderings reduce problem size. RCL-OG first defines reinforcement learning with new reward mechanism in an efficacy way, deep Q-learning update transfer rewards between nodes. Next, obtains probability transition nodes computes orderings. can sample obtain ordering high probability. Experiments synthetic benchmark datasets show provides accurate approximation achieves better results than competitive discovery algorithms.