1: \begin{abstract}

2: Machine learning for partial differential equations (PDEs)

3: is a hot topic. In this paper we introduce and analyse a Deep BSDE scheme for

4: nonlinear integro-PDEs with unbounded nonlocal operators -- problems arising in e.g.~stochastic control and games involving infinite activity jump-processes. The scheme is based on a stochastic forward-backward SDE representation of the solution of the PDE and (i) approximation of small jumps by a Gaussian process, (ii) simulation of the forward part, and (iii) a neural net regression for the backward part. Unlike grid-based schemes, it does not suffer from the curse of dimensionality and is therefore suitable for high dimensional problems. The scheme is designed to be convergent even in the infinite activity/unbounded nonlocal operator case. A full convergence analysis is given and constitutes the main part of the paper.

5: % Machine learning techniques for partial differential equations (PDEs)

6: % is a hot topics, and different algorithms proposed in the literature show efficient numerical approximation in high dimension. In this article, we introduce and study a deep learning based approximation algorithm for solutions of  high dimensional nonlocal nonlinear partial differential equations. We give an efficient approximation of the nonlocal term without neglecting the small jumps. Our methodologyD extends the approach recently proposed in \cite{HPW20} for semi-linear PDEs. We prove the convergence of the proposed algorithm.

7: \end{abstract}

8: