Researchers unveil PENN-GMD for mixed-noise stochastic systems
Neural network PENN-GMD cracks 5 complex stochastic systems with mixed-noise dynamics...
A team led by Xiaolong Wang from [Shenzhen University] has developed PENN-GMD (Parameter Estimation Neural Network with Gaussian Mixture Distribution), a novel approach to tackle parameter identification in stochastic dynamical systems driven by mixed noises—environments where traditional likelihood-based methods often stumble due to intractable likelihood functions.
PENN-GMD leverages a neural network architecture that outputs a full-covariance Gaussian Mixture Distribution over system parameters, allowing it to explicitly model parameter couplings and multi-modal likelihood structures. Unlike conventional methods, it uses a surjective parameterization that hard-codes all GMD constraints during training, minimizing negative log-likelihood to approximate true likelihoods. The team validated PENN-GMD on five progressively complex systems: fractional Gaussian noise, Lévy processes, colored noise oscillators, coupled neuron models with varying observability, and an aeroelastic airfoil with unidentifiable stochastic disturbances. Results show PENN-GMD accurately recovers likelihood distributions, captures parameter interactions, and identifies non-identifiability through variance broadening or mode splitting—capabilities previously infeasible with conventional tools.
- PENN-GMD maps partially observed trajectories to a full-covariance Gaussian Mixture Distribution (GMD) to model parameter couplings in mixed-noise stochastic systems
- Validated on 5 complex benchmarks, including fractional Gaussian, Lévy, and colored noise systems, and an aeroelastic airfoil model with unidentifiable disturbances
- Diagnoses non-identifiability via variance broadening or mode splitting, outperforming traditional likelihood-based methods
Why It Matters
PENN-GMD enables robust parameter estimation in noisy, real-world systems where traditional methods fail—critical for climate modeling, neuroscience, and aerospace engineering.