DeepMartNet -- A Martingale based Deep Neural Network Learning Algorithm for Eigenvalue/BVP Problems and Optimal Stochastic Controls

TL;DR

DeepMartNet introduces a neural network approach leveraging martingale properties to efficiently solve high-dimensional eigenvalue, boundary value, and stochastic control problems with various boundary conditions.

Contribution

It presents a novel martingale-based loss function for neural networks to solve complex PDEs and stochastic control problems in high dimensions.

Findings

Effective in high-dimensional settings

Handles various boundary conditions

Applicable to stochastic control problems

Abstract

In this paper, we propose a neural network learning algorithm for solving eigenvalue problems and boundary value problems (BVPs) for elliptic operators and initial BVPs (IBVPs) of quasi-linear parabolic equations in high dimensions as well as optimal stochastic controls. The method is based on the Martingale property in the stochastic representation for the eigenvalue/BVP/IBVP problems and martingale principle for optimal stochastic controls. A loss function based on the Martingale property can be used for efficient optimization by sampling the stochastic processes associated with the elliptic operators or value process for stochastic controls. The proposed algorithm can be used for eigenvalue problems and BVPs and IBVPs with Dirichlet, Neumann, and Robin boundaries in bounded or unbounded domains and some feedback stochastic control problems.