Solving McKean-Vlasov Equation by deep learning particle method

TL;DR

This paper presents a deep learning-based meshless method for solving McKean-Vlasov SDEs, overcoming limitations of traditional particle methods by using PINNs to improve efficiency and scalability for complex, high-dimensional problems.

Contribution

The paper introduces a novel PINNs-based meshless solver for MV-SDEs that does not depend on the propagation of chaos, enabling more efficient simulations over long time horizons.

Findings

Demonstrates improved accuracy over traditional methods

Shows scalability to high-dimensional problems

Provides theoretical error estimates for the loss function

Abstract

We introduce a novel meshless simulation method for the McKean-Vlasov Stochastic Differential Equation (MV-SDE) utilizing deep learning, applicable to both self-interaction and interaction scenarios. Traditionally, numerical methods for this equation rely on the interacting particle method combined with techniques based on the It\^o-Taylor expansion. The convergence rate of this approach is determined by two parameters: the number of particles $N$ and the time step size $h$ for each Euler iteration. However, for extended time horizons or equations with larger Lipschitz coefficients, this method is often limited, as it requires a significant increase in Euler iterations to achieve the desired precision $\epsilon$. To overcome the challenges posed by the difficulty of parallelizing the simulation of continuous interacting particle systems, which involve solving high-dimensional coupled SDEs, we propose a meshless MV-SDE solver grounded in Physics-Informed Neural Networks (PINNs) that does not rely on the propagation of chaos result. Our method constructs a pseudo MV-SDE using It\^o calculus, then quantifies the discrepancy between this equation and the original MV-SDE, with the error minimized through a loss function. This loss is controlled via an optimization algorithm, independent of the time step size, and we provide an error estimate for the loss function. The advantages of our approach are demonstrated through corresponding simulations.