1: \begin{abstract}%   <- trailing '%' for backward compatibility of .sty file

2:   This paper considers the robust and efficient implementation of

3:   Gaussian process regression with a Student-$t$ observation model.

4:   The challenge with the Student-$t$ model is the analytically

5:   intractable inference which is why several approximative methods

6:   have been proposed. The expectation propagation (EP) has been

7:   found to be a very accurate method in many empirical studies but

8:   the convergence of the EP is known to be problematic with models

9:   containing non-log-concave site functions such as the Student-$t$

10:   distribution. In this paper we illustrate the situations where

11:   the standard EP fails to converge and review different

12:   modifications and alternative algorithms for improving the

13:   convergence. We demonstrate that convergence problems may occur

14:   during the type-II maximum a posteriori (MAP) estimation of the

15:   hyperparameters and show that the standard EP may not converge in

16:   the MAP values in some difficult cases. We present a robust

17:   implementation which relies primarily on parallel EP updates and

18:   utilizes a moment-matching-based double-loop algorithm with

19:   adaptively selected step size in difficult cases. The predictive

20:   performance of the EP is compared to the Laplace, variational

21:   Bayes, and Markov chain Monte Carlo approximations.

22: %

23: %According to our knowledge no numerically robust implementation of the Student-$t$

24: %model has been presented.

25: %

26: %

27: %with artificial regression examples

28: %

29: \end{abstract}

30: