Augmented Physics-Based Models for High-Order Markov Filtering

Shuo Tang, Tales Imbiriba, Jindřich Duník, Ondřej Straka, Pau Closas

TL;DR

This paper introduces two novel methods, AG-APBM and AP-APBM, to improve state estimation in high-order Markov models, even when the exact system dynamics or Markov order are unknown. These methods extend augmented physics-based models (APBMs) by incorporating past states into the model, offering better accuracy than standard APBMs while AP-APBM significantly reduces computational cost.

What Problem Does This Paper Address?

Accurate state estimation is crucial in many engineering applications, especially when dealing with noisy measurements and complex system dynamics. Traditional methods like Kalman filters work well for linear systems, but nonlinear systems require more advanced techniques. Hybrid approaches, combining physics-based models (PBMs) with data-driven machine learning (ML) models (known as Augmented Physics-Based Models or APBMs), offer a way to leverage existing physical knowledge while allowing ML to learn unmodeled dynamics.

However, existing APBMs primarily focus on first-order Markov processes. Many real-world systems exhibit higher-order Markovian behavior, meaning their current state depends on multiple past states, not just the immediate previous one. While techniques like state augmentation (AGS) can handle high-order Markov models, they typically require complete knowledge of the system dynamics and the Markov order. This paper addresses the challenge of state estimation for high-order Markov models when system dynamics are not fully known or the Markov order is uncertain.

Methods

The core contribution is the extension of the APBM framework to handle high-order Markov models. Two main approaches are proposed:

  • Augmented-State APBM (AG-APBM): This method directly augments the state vector to include past states, effectively transforming the high-order Markov model into a first-order one. This allows the APBM to learn the dynamics of this augmented state. The PBM component is combined with a neural network (NN) that learns the deviations from the PBM.
  • Approximated-State APBM (AP-APBM): To mitigate the computational burden of the AG-APBM's high-dimensional state space, this approach uses point estimates of past states (e.g., their expected values) as input to the NN, rather than the full posterior distributions. This approximation significantly reduces computational cost while aiming to maintain high accuracy.

Both methods leverage a hybrid approach where a physics-based model (PBM) provides a baseline, and a neural network learns to correct or augment it. The authors use a Cubature Kalman Filter (CKF) for state estimation, which is suitable for nonlinear systems.

Datasets

The paper does not introduce new datasets. Instead, it uses simulated data generated from:

  • An Autoregressive (AR(3)) model: This is a standard high-order Markov model used to evaluate the filtering performance in a controlled environment.
  • A nonlinear delayed-feedback control model: This simulates a target tracking scenario with communication delays, representing a more realistic and complex application.

Key Insights and Ideas

  • Bridging the Gap for High-Order Markov Models: The paper successfully extends the APBM framework to high-order Markov models, a significant step beyond existing APBM literature.
  • Balancing Accuracy and Computational Cost: The AP-APBM offers a compelling trade-off, achieving performance close to the more computationally intensive AG-APBM while drastically reducing time and memory usage. This is crucial for real-time applications.
  • Learning with Partial Knowledge: The methods demonstrate effectiveness even when the exact Markov order or full system dynamics are unknown, highlighting the robustness of the hybrid approach.
  • NN as a "Corrector": The NN in the APBM framework acts as a powerful tool to learn and compensate for the limitations of the physics-based model, especially in complex, high-order dynamics.

Key Results

  • Improved Accuracy: Both AG-APBM and AP-APBM significantly outperformed the standard PBM-only model and a first-order APBM in both the AR(3) model and the target tracking scenarios.
  • In the AR(3) model, AG-APBM and AP-APBM reduced estimation error by 31.1% and 26.7%, respectively, compared to the PBM.
  • In the target tracking scenario, AG-APBM and AP-APBM reduced the error by 58.8% and 52.9% compared to the PBM.
  • Computational Efficiency: AP-APBM showed substantial computational savings compared to AG-APBM.
  • Time cost was reduced by 37.5%.
  • Memory usage was reduced by 20%.
  • AGS Benchmark: The proposed methods achieved RMSEs comparable to the Augmented State (AGS) approach, which assumes perfect knowledge of the system dynamics and Markov order, demonstrating their effectiveness.

Limitations and Future Work

  • Unknown Noise Parameters: The current work assumes known noise statistics. Future research could focus on dealing with unknown noise parameters, potentially using data-driven methods to estimate them.
  • Non-Gaussian Distributions: The current APBM framework is primarily applied to Gaussian noise assumptions. Extending it to handle non-Gaussian state-space distributions would broaden its applicability.
  • Alternative NN Architectures: The paper used Multilayer Perceptrons (MLPs). Exploring other NN architectures like Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), or Transformers could potentially offer better performance for learning complex, long-term dependencies in high-order Markov dynamics.
  • Approximation Trade-offs: While AP-APBM offers significant computational benefits, the paper notes a slight degradation in accuracy compared to AG-APBM. Further research could explore ways to minimize this accuracy gap.

Summary updated

BibTeX
@misc{tang2024augmented,
  author = {Shuo Tang and Tales Imbiriba and Jindřich Duník and Ondřej Straka and Pau Closas},
  title = {{Augmented Physics-Based Models for High-Order Markov Filtering}},
  year = {2024},
  doi = {10.2139/ssrn.4681699},
  howpublished = {SSRN},
}
Figure