Three algorithms for solving high-dimensional fully-coupled FBSDEs through deep learning

Shaolin Ji, Shige Peng, Ying Peng, Xichuan ZhangIEEE Intell. Syst 2019 · OpenReview

TL;DR

This paper introduces three novel deep learning algorithms to solve complex, high-dimensional forward-backward stochastic differential equations (FBSDEs). These algorithms offer a powerful new tool for tackling problems in finance and AI that were previously computationally intractable due to the "curse of dimensionality."

What Problem Does This Paper Address?

Forward-backward stochastic differential equations (FBSDEs) are a powerful mathematical framework used in various fields, including financial modeling (e.g., for derivatives pricing with credit and funding adjustments) and AI applications like robotics. However, finding analytical solutions to FBSDEs is often impossible, especially when dealing with high-dimensional systems. Traditional numerical methods struggle with these high-dimensional problems due to the "curse of dimensionality," where computational complexity grows exponentially with the number of dimensions. This paper aims to overcome this challenge by leveraging deep learning to develop efficient and accurate numerical solutions for fully coupled, high-dimensional FBSDEs.

Methods

The authors adapt an optimal control perspective to solve FBSDEs, viewing the problem as finding optimal control strategies for neural networks. They propose three distinct algorithms, each employing a different feedback control strategy:

  • Algorithm 1 (Feedback based on X, Y): This method extends existing work by treating the Z term as a feedback control dependent on the state processes X and Y. It's generalized to handle cases where the drift and diffusion coefficients depend on Z.
  • Algorithm 2 (Feedback based on X): This approach introduces a new control variable 'u' for the Y process in the forward SDE, making both 'u' and Z feedback controls dependent solely on the state X. This requires modifying the cost functional.
  • Algorithm 3 (Feedback based on X, Y, Z with Picard Iteration): Inspired by Picard iteration, this method uses the Z term from the previous iteration as input for the current iteration, effectively incorporating feedback from X, Y, and Z. This method requires specific assumptions for convergence.

All algorithms utilize fully connected neural networks with ReLU activation and Adam optimization. The core idea is to train these networks to approximate the unknown functions within the FBSDE system.

Datasets

Not discussed in the paper. The paper focuses on solving a general class of FBSDEs and demonstrates its methods with synthetic examples, not on training with real-world datasets.

Key Insights and Ideas

  • Deep Learning as an Optimal Control Framework: The paper effectively reframes solving FBSDEs as an optimal control problem, where neural network parameters act as controls and a defined loss function serves as the optimization objective.
  • Handling High-Dimensionality: The proposed deep learning approaches demonstrate remarkable performance in high-dimensional settings, offering a viable solution to the "curse of dimensionality" that plagues traditional numerical methods.
  • Flexible Feedback Strategies: The three algorithms offer different levels of feedback complexity (X,Y; X; X,Y,Z), allowing for tailored solutions depending on the problem's structure and desired accuracy.
  • Algorithm 3's Stability: While potentially more computationally intensive per iteration, Algorithm 3, leveraging Picard iteration, shows superior accuracy and stability in achieving convergence.

Key Results

The paper presents numerical results for various FBSDE examples, including partially-coupled and fully-coupled cases, and across different dimensions (d=1, d=100).

  • High Accuracy: For high-dimensional cases (d=100), the algorithms achieve relative errors as low as 0.1% for the initial value of Y0.
  • Remarkable Performance: The methods demonstrate significant improvements over existing approaches, particularly for high-dimensional problems where other methods are not applicable.
  • Algorithm Comparison: Algorithm 3 generally shows better accuracy and stability, though it may require more time per iteration. Algorithm 2 is fast but can struggle with stable convergence. Algorithm 1 offers a balance.
  • Convergence: All algorithms demonstrate convergence, with the loss function decreasing significantly over iterations.

Limitations and Future Work

  • Assumption Requirements: Algorithm 3, in particular, relies on specific assumptions for convergence, which might not always be met. The paper notes that it handles situations with small T and Lipschitz coefficients well.
  • Computational Cost: While effective, deep learning methods can be computationally intensive, especially for Algorithm 3 which requires more time per iteration.
  • Future Work: The authors mention establishing concrete assumptions for the feedback function $\phi$ to ensure discrete-time scheme convergence. They also suggest that their methods can handle situations where the standard assumptions are not met. The paper implies that further exploration of these algorithms for FBSDEs with random coefficients is a potential future direction.

Summary updated

Abstract

Recently, the deep learning method has been used for solving forward-backward stochastic differential equations (FBSDEs) and parabolic partial differential equations (PDEs). It has good accuracy and performance for high-dimensional problems. In this paper, we mainly solve fully coupled FBSDEs through deep learning and provide three algorithms. Several numerical results show remarkable performance especially for high-dimensional cases.
BibTeX
@inproceedings{ji2019three,
  author = {Shaolin Ji and Shige Peng and Ying Peng and Xichuan Zhang},
  title = {{Three algorithms for solving high-dimensional fully-coupled FBSDEs through deep learning}},
  booktitle = {IEEE Intell. Syst},
  year = {2019},
  url = {https://openreview.net/forum?id=LKy4KpdWaA},
}
Figure