1: \begin{abstract}

2: In this paper, we propose forward and backward stochastic differential

3: equations (FBSDEs) based deep neural network (DNN) learning algorithms for the

4: solution of high dimensional quasilinear parabolic partial differential

5: equations (PDEs), which are related to the FBSDEs by the Pardoux-Peng theory. The algorithms rely on a learning process by minimizing

6: the pathwise difference between two discrete stochastic processes,

7: defined by the time discretization of the FBSDEs and the DNN representation of

8: the PDE solutions, respectively. The proposed algorithms are shown to generate DNN solutions for a  100-dimensional Black--Scholes--Barenblatt equation,

9: accurate in a finite region in the solution space, and has a convergence rate similar to that of the Euler--Maruyama discretization used for the FBSDEs. As a result, a Richardson extrapolation technique over time discretizations can be used to enhance the accuracy of the DNN solutions.

10: For time oscillatory solutions, a multiscale DNN is shown to improve the performance of the FBSDE DNN for high frequencies.

11: 

12: \end{abstract}

13: