The deep multi-FBSDE method: a robust deep learning method for coupled FBSDEs
Kristoffer Andersson, Adam Andersson, Cornelis W. Oosterlee · 2025
Read The deep multi-FBSDE method: a robust deep learning method for coupled FBSDEs by Kristoffer Andersson, Adam Andersson, Cornelis W. Oosterlee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-13193.txt.
Introduction
Kristoffer Andersson, Adam Andersson, and Cornelis Oosterlee give a deep method for coupled forward-backward stochastic equations. Their paper is from two thousand twenty-five; the pack row gives no title or year, so the text head is the source. Such equation pairs show up in control and finance. A known deep method solves many, but fails to settle in some coupled cases.
Their fix works in two phases. First they fit the start value of a whole family of twin problems that share the same linked partial equation. Then they solve the true problem from that start. Tests show their method settles where the standard one does not, which widens the reach of deep solvers for this class.
Catalog record
| CREATOR | Kristoffer Andersson, Adam Andersson, Cornelis W. Oosterlee |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-13193
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The deep multi-FBSDE method: a robust deep learning method for coupled FBSDEs Kristoffer Andersson, Adam Andersson, Cornelis W. Oosterlee Abstract We introduce the deep multi-FBSDE method for robust approximation of coupled forward– backward stochastic differential equations (FBSDEs), focusing on cases where the deep BSDE method of Han, Jentzen, and E (2018) fails to converge. To overcome the convergence issues, we consider a family of FBSDEs that are equivalent to the original problem in the sense that they satisfy the same associated partial differential equation (PDE) and initial value. Our algorithm proceeds in two phases: first, we approximate the initial condition jointly for a small number of FBSDEs from the FBSDE family, and second, we approximate the original FBSDE using the initial condition approximated in the first phase. Numerical experiments show that our method converges even when the standard deep FBSDE method does not. 1 Introduction Solving forward-backward stochastic differential equations (FBSDEs) using deep learning has emerged as a powerful technique to overcome the curse of dimensionality inherent in classical numerical methods. Han, Jentzen, and E’s seminal work [ 20 ] introduced the deep BSDE method , employing neural networks to approximate solutions of high-dimensional parabolic partial differential equations (PDEs) by reformulating them as FBSDEs. This breakthrough demonstrated remarkable performance on problems previously considered intractable, showcasing the ability of neural networks to efficiently handle complex nonlinearities at scales unattainable by traditional approaches. Following this success, several deep learning-based strategies have emerged, notably [ 6 , 7 , 5 , 17 , 31 , 26 , 27 , 34 , 22 , 3 ] , with convergence analyses provided in, e.g., [ 21 , 24 , 9 , 15 , 19 , 25 , 32 , 28 , 2 ] . Concurrently, a separate branch known as backward-type methods, closer in spirit to classical dynamic programming algorithms, has developed, see e.g., [ 23 , 14 , 16 , 30 , 29 , 23 , 18 ] . For a comprehensive overview of machine learning algorithms for PDE approximation, we refer to [ 8 ] . Despite these advancements, forward-type deep BSDE methods, including the original method in [ 20 ] , still face significant convergence challenges when addressing certain classes of FBSDEs, such as those encountered in stochastic control problems [ 1 ] and beyond. Specifically, these forward-type approaches often fail to converge or become trapped in suboptimal solutions due to ill-conditioned optimization landscapes [ 23 , 1 , 10 ] . Although backward-type methods may not suffer from precisely the same convergence issues, they encounter different challenges, especially implementation challenges when dealing with coupled FBSDEs (see [ 12 , 11 , 13 ] for further discussions of implementation challenges for backward type methods when dealing with coupled FBSDEs) and accumulated errors in general (see e.g., [ 30 ] ). For FBSDEs related to stochastic control problems through dynamic programming, a robust deep FBSDE method was proposed in our paper [ 1 ] . The scope of the present paper is the introduction of another robust modification of the deep FBSDE method, the deep multi-FBSDE method . It is applicable to a broader class of FBSDEs and their PDEs, not strictly defined due to the experimental nature of our paper, but not restricted to only control problems. By robustness, we mean that the method can be optimized for the equation of interest, and not the other way around, which is sometimes seen in the deep learning-based scientific computing literature. Of course, this is an inprecise definition and no method allows convergence for any equation. Still, our method has been working on the vast majority of equations we have tested and we have deliberately chosen challenging equations. We refer to the original deep BSDE method, applied directly to coupled FBSDE without modification, as the deep FBSDE method [ 1 , 26 ] . Our proposed approach alleviates optimization difficulties by utilizing a family of FBSDEs that share the same initial condition as well as the same functional forms as the original FBSDE. In other words, all FBSDEs in this family are solved by the same PDE. We propose a two-phase optimization procedure. In the first phase, we approximate the joint initial conditions for all FBSDEs. In the second phase, we apply a reduced version of the deep FBSDE method, where we optimize only over the control processes while using the initial condition obtained from the first phase. This leads to a numerical scheme which we experimentally show has consistent and accurate convergence, even for strongly coupled, nonlinear FBSDEs. Our numerical experiments demonstrate stable and accurate solutions for problems for which the classical deep FBSDE methods fail. In summary, the primary contributions of this paper are as follows. • A robust deep FBSDE method that we empirically demonstrate solves the convergence issues faced by existing deep learning-based methods for strongly coupled FBSDEs. • A general purpose training strategy that is demonstrated numerically to be applicable to general FBSDE systems, without restrictive assumptions such as the existence of an equivalent stochastic control problem, as in [ 1 ] . • Comprehensive numerical demonstrations illustrating the effectiveness and robustness of our method, especially in challenging scenarios involving stochastic control and nonlinear PDEs, thus underscoring its broad applicability and efficiency. The rest of this paper is organized as follows: Section 2 presents the FBSDE problem and its solution by the non-linear Feynman–Kac formula through the PDE. Section 3 introduces our proposed robust deep learning method, detailing its formulation and implementation. Numerical results demonstrating its advantages are presented in Section 4, and we end in Section 5 with conclusions and future directions. 2 Forward backward stochastic differential equations Throughout this paper, we let T\in(0,\infty) , d,k\in\mathbb{N} , x_{0}\in\mathbb{R}^{d} , (W_{t})_{t\in[0,T]} be a k -dimensional standard Brownian motion on a filtered probability space (\Omega,\mathcal{F},(\mathcal{F}_{t})_{t\in[0,T]},\mathbb{P}) and the problem coefficients be denoted by b\colon[0,T]\times\mathbb{R}^{d}\times\mathbb{R}\times\mathbb{R}^{k}\to\mathbb{R}^{d} , \sigma\colon[0,T]\times\mathbb{R}^{d}\to\mathbb{R}^{d\times k} , g\colon\mathbb{R}^{d}\to\mathbb{R} and f\colon[0,T]\times\mathbb{R}^{d}\times\mathbb{R}\times\mathbb{R}^{k}\to\mathbb{R} . The general form of the FBSDEs considered is given by the coupled system of equations \begin{cases}X_{t}=x_{0}+\int_{0}^{t}b(s,X_{s},Y_{s},Z_{s})\text{d}s+\int_{0}^{t}\sigma(s,X_{s})\text{d}W_{s},\\ Y_{t}=g(X_{T})+\int_{t}^{T}f(s,X_{s},Y_{s},Z_{s})\text{d}s-\int_{t}^{T}\langle Z_{s},\sigma(s,X_{s})\text{d}W_{s}\rangle,\quad t\in[0,T].\end{cases} (1) The associate PDE is then a semilinear parabolic PDE on the form \begin{cases}\frac{\partial v}{\partial t}(t,x)+\mathcal{L}v(t,x)+f\big(t,x,v(t,x),\mathrm{D}_{x}v(t,x)\sigma(t,x)\big)=0,&(t,x)\in[0,T)\times\mathbb{R}^{d},\\ v(T,x)=g(x),&x\in\mathbb{R}^{d}.\end{cases} (2) Here \mathcal{L} is the infinitesimal generator, acting on a twice continuously differentiable functions \rho by: \displaystyle\mathcal{L}\rho(x)= \displaystyle b\big(t,x,\rho(t,x),\mathrm{D}_{x}\rho(t,x)\sigma(t,x)\big)^{\top}\mathrm{D}_{x}\rho(x)+\frac{1}{2}\text{Tr}\big(a(t,x)\mathrm{D}_{x}^{2}\rho(x)\big) \displaystyle= \displaystyle\sum_{i=1}^{d}b_{i}\big(t,x,\rho(t,x),\mathrm{D}_{x}\rho(t,x)\sigma(t,x)\big)\partial_{i}\rho(x)+\frac{1}{2}\sum_{i,j=1}^{d}a_{ij}(t,x)\partial_{i,j}\rho(x), where a=\sigma\sigma^{\top} and \text{Tr}(\cdot) denotes the trace of a matrix. The connection between ( 1 ) and ( 2 ) is established through the nonlinear Feynman–Kac formula, see e.g., [ 35 ] , which states that Y_{t}=v(t,X_{t}),\ \text{and}\ Z_{t}=\mathrm{D}_{x}v(t,X_{t}). Note that in the above, we have implicitly assumed that v is differentiable with respect to the spatial variable. Although this holds true when ( 2 ) possesses a classical solution, we can often only expect a solution in the viscosity sense. In this situation, derivatives of v with respect to the space variable should be interpreted as set-valued sub-derivatives. Nevertheless, from our perspective, which is to find approximations of ( 1 ) and ( 2 ), we can overlook this obstacle and view Z as defined above. 3 The deep FBSDE and deep multi-FBSDE methods In this section, we first introduce the classical deep FBSDE method and briefly discuss its convergence problems for coupled FBSDEs. Most importantly, we introduce our novel modification