1: \begin{abstract}

2:   The pricing of Bermudan options amounts to solving a dynamic programming

3:   principle, in which the main difficulty, especially in high dimension, comes

4:   from the conditional expectation involved in the computation of the

5:   continuation value. These conditional expectations are classically computed by

6:   regression techniques on a finite dimensional vector space. In this work, we

7:   study neural networks approximations of conditional expectations. We prove the

8:   convergence of the well-known Longstaff and Schwartz algorithm when the

9:   standard least-square regression is replaced by a neural network

10:   approximation. We illustrate the numerical efficiency of neural networks as an

11:   alternative to standard regression methods for approximating conditional

12:   expectations on several numerical examples.

13: 

14:   \noindent\textbf{Key words}: Bermudan options, optimal stopping,

15:   regression methods, deep learning, neural networks.

16: 

17:   % \noindent\textbf{AMS subject classification}: 62L20, 62L15, 91G60, 65Y05, 60H07

18: \end{abstract}

19: