Improving greedy local search methods by switching the search space

Bayesian networks play a vital role in human understanding of the world. Finding a precise equivalence class of a Bayesian network is an effective way to represent causality. However, as one of the most widely used methods of searching for equivalence classes, greedy equivalence search (GES), can easily fall into a local optimum. To address this problem, we explore the reasons why GES becomes stuck in a local optimum by analyzing its operators and search strategies in detail. Moreover, we demonstrate that converting the search space into another space can address the drawbacks of local search in the space of the equivalence class. Accordingly, two novel frameworks based on switching spaces are proposed to improve GES. Finally, the effectiveness, scalability, and stability of the proposed methods are verified by extensive experiments through which our frameworks are compared with state-of-the-art methods on different benchmarks. The results show that our methods significantly strengthen the performance of GES.