Abstract
Stochastic dynamical systems provide essential mathematical frameworks for modeling complex real-world phenomena. The Fokker-Planck-Kolmogorov (FPK) equation governs the evolution of probability density functions associated with stochastic system trajectories. Developing robust numerical methods for solving the FPK equation is critical for understanding and predicting stochastic behavior. Here, we introduce the distribution self-adaptive normalized physics-informed neural network (DSN-PINNs) for solving time-dependent FPK equations through the integration of soft normalization constraints with adaptive resampling strategies. Specifically, we employ a normalization-enhanced PINN model in a pretraining phase to establish the solution's global structure and scale, generating a reliable prior distribution. Subsequently, guided by this prior, we dynamically reallocate training points via weighted kernel density estimation, concentrating computational resources on regions most representative of the underlying probability distribution throughout the learning process. The key innovation lies in our method's ability to exploit the intrinsic structural properties of stochastic dynamics while maintaining computational accuracy and implementation simplicity. We demonstrate the framework's effectiveness through comprehensive numerical experiments and comparative analyses with existing methods, including validation on real-world economic datasets.
The Challenge of Stochastic Dynamics
Stochastic dynamical systems are ubiquitous, appearing in fields ranging from financial market fluctuations to chemical diffusion. The Fokker-Planck-Kolmogorov (FPK) equation is the fundamental partial differential equation (PDE) used to describe how the probability density of such systems evolves over time. While Physics-Informed Neural Networks (PINNs) have emerged as a powerful mesh-free tool for solving PDEs, they face a significant hurdle with the FPK equation: the "trivial solution problem."
Because FPK boundary conditions typically state that probability vanishes at infinity, standard PINNs often converge to a zero solution. This mathematically satisfies the PDE and boundary conditions but fails to represent a physical probability distribution. To overcome this, the total probability must be constrained to equal one (normalization), a task that becomes computationally prohibitive in high-dimensional spaces using traditional integration methods.
The DSN-PINN Framework
To address these limitations, the Distribution Self-adaptation Normalized PINN (DSN-PINN) framework introduces a two-stage approach that ensures the network respects the laws of probability while focusing computational effort where it matters most.
Normalization Design (N-PINNs)
Instead of rigid grid-based integration, DSN-PINNs implement a "soft" normalization constraint. The model calculates the average value of the network's output over the existing sample points and penalizes deviations from the expected scale. This approach keeps computational costs linear rather than exponential, making it viable for complex systems.
Distribution Self-adaptation (D-PINNs)
Following an initial "warm start" phase where the model learns the basic global structure, the algorithm performs adaptive resampling. Using weighted kernel density estimation (KDE), the model identifies regions of high probability mass and reallocates training points to these areas.
To maintain stability and prevent overfitting to high-density peaks, the framework employs a mixture strategy. It combines a percentage ($\beta$) of resampled points with a set of original uniform "base" points ($(1-\beta)$). This ensures the model captures both the sharp features of the distribution and the behavior in the tails.
Experimental Validation
The researchers evaluated DSN-PINNs across several benchmarks, including 1D and 2D nonlinear oscillators and a real-world application involving Chinese A-share market data (2019–2022). In the financial application, the dynamics of 4,500 stocks were modeled as Geometric Brownian Motion.
Key results include:
- Enhanced Accuracy: In synthetic benchmarks, DSN-PINNs achieved a Maximum Absolute Error (MAE) of 0.0048, significantly lower than the 0.0112 recorded for standard PINNs.
- Stability in Real-World Data: The model maintained a low Mean Squared Error (MSE) of $2.03 \times 10^{-5}$ over a two-year simulated trading period, successfully capturing the log-normal distribution of stock prices where standard PINNs failed.
- Physics Consistency: The Mean PDE Residual was often an order of magnitude lower than competing methods, indicating a more faithful adherence to the underlying stochastic physics.
Conclusion and Future Directions
DSN-PINNs represent a significant advancement in making neural networks "probability-aware." By combining soft normalization with dynamic resource allocation, the framework avoids the trivial solution trap and handles the sharp features typical of stochastic systems. Future work will focus on extending the framework to handle non-local operators (such as Lévy noise) and further scaling the method for extremely high-dimensional problems where sampling density remains a challenge.
Summary updated