1: \begin{abstract}

2: Probabilistic graphical models are a key tool in machine learning applications. Computing the partition function, i.e., normalizing constant,

3: is a fundamental task of statistical inference but it is generally computationally intractable, leading to extensive study of

4: %Various

5: approximation methods. % have been extensively studied in the literature

6: %to resolve the problem, where

7: Iterative variational methods %schemes of variational type

8: are a popular and successful family of approaches. % in practice.

9: However, even state of the art variational methods

10: can return poor results or fail to converge on difficult instances. %, thus

11: %provide extremely bad estimations.

12: %This paper follows an alternative route of computing

13: In this paper,

14: we instead consider computing the partition function via sequential summation over %(elimination of)

15: variables. We develop robust approximate algorithms by combining ideas from

16: mini-bucket elimination %process from machine learning

17: with tensor network and renormalization group methods from statistical physics.

18: The resulting ``convergence-free'' methods show good

19: empirical performance on both synthetic and real-world benchmark models,

20: %approximation guarantees

21: even for difficult instances.

22: % can be varied . % efficiently.

23: %Our experimental results confirm the theoretical prediction: the proposed bucket elimination algorithms outperform other state-of-the-art elimination and variational algorithms on both synthetic and real-world benchmark models.

24: \end{abstract}

25: