Efficient Generative Adversarial DAG Learning with No-Curl

Causal structure learning from data is a challenging task as the search space is typically huge. In recent years, a series of methods have been proposed to reformulate causality learning into an optimization problem with a continuous acyclicity constraint to allow problem-solving with continuous optimization techniques. This paper further improves on the causality learning results and efficiency of the continuous optimization approach through the use of generative adversarial neural network learning, which overcomes the limitations of using maximum likelihood estimation in the existing methods. In addition, we adapt the recently proposed DAG-NoCurl framework to the generative causal structure learning to improve speed performance. In particular, our adapted method does not constrain causal structure discovery to its initial estimation, hence allowing further improvement of the learning results. The proposed method has been tested on several benchmarks to compare against the state-of-the-art.