Parameter estimation in differential equation models with constrained states

David A. Campbell, Giles Hooker, Kimberley B. McAuleyJournal of Chemometrics 2012 · DOI

TL;DR

This paper introduces a novel method for estimating parameters and states in differential equation models, specifically addressing the challenge of ensuring that estimated system states are physically meaningful (e.g., non-negative concentrations). By extending generalized smoothing techniques with interpretability constraints, the authors provide a more robust approach to modeling complex systems, particularly in chemical engineering, where previous methods struggled with these constraints.

What Problem Does This Paper Address?

Modeling dynamic systems often involves using differential equations (ODEs) to describe how system states change over time, influenced by external inputs and model parameters. A key challenge in this process is parameter estimation – determining the values of these model parameters from observed data. However, many real-world systems have inherent constraints on their states. For instance, chemical concentrations cannot be negative. Traditional parameter estimation methods, like nonlinear least squares (NLS) or Markov Chain Monte Carlo (MCMC), often struggle when these constraints are present. They might produce physically impossible results (e.g., negative concentrations) or require strong, potentially biasing, assumptions about initial conditions. Existing methods that attempt to enforce these constraints can either limit the model's flexibility, leading to poor fits, or fail to adequately address the problem, especially when dealing with complex dynamics or unobserved states. This paper aims to bridge this gap by developing a method that estimates both model parameters and system states while explicitly enforcing these crucial interpretability constraints.

Methods

The core of the paper's contribution lies in extending the Generalized Smoothing (GS) method. GS is a statistical technique that estimates system states and model parameters by balancing two competing objectives: fitting the observed data and adhering to the underlying differential equation model. This balance is controlled by a smoothing parameter, $\lambda$.

The novel aspects introduced are:

  • Interpretability Constraints: The authors incorporate functions, denoted as $g(\cdot)$, into the smoothing process. These functions act as "interpretable smoothers," ensuring that the estimated system states conform to desired properties, such as non-negativity or monotonicity. This is achieved by modifying the basis expansion of the system states, $x(t) = g(b'\phi(t))$, where $b$ are basis coefficients.
  • Profile Likelihood Estimation: To estimate the ODE parameters ($\theta$), the method uses a profile likelihood approach. This involves optimizing for the basis coefficients ($b$) at each step of the $\theta$ optimization, effectively "profiling out" the nuisance parameters related to the smooth.
  • Intuitive $\lambda$ Estimation Strategy: A key challenge in GS is selecting the smoothing parameter $\lambda$. The paper proposes an iterative annealing strategy: starting with a small $\lambda$ (low bias, high variance) and gradually increasing it. This process helps to refine parameter estimates and avoid local optima, while also providing a way to balance model fidelity and data fitting.
  • Handling Unobserved States: The method is designed to estimate unobserved system states by regularizing them through their impact on the fit to observed states, while still penalizing them to follow the ODE dynamics.

Datasets

The primary dataset used for evaluation is from a real-world chemical engineering example: the nylon production system. This dataset involves multiple experimental runs with step changes in input conditions and noisy, sometimes unevenly sampled, observations of chemical concentrations (A, C, and W). The authors also utilize simulated data generated from variations of the nylon model to explore the behavior of their method under different conditions, including model misspecification.

Key Insights and Ideas

  • Constraints are Crucial for Realism: The paper strongly emphasizes that for many physical systems, enforcing interpretability constraints (like non-negativity) on estimated states is not just a nicety but a necessity for obtaining physically meaningful and reliable parameter estimates.
  • Balancing Act: Model vs. Data: The core of the GS method, and its extension here, is the elegant way it balances fitting noisy data with adhering to a theoretical model. The smoothing parameter $\lambda$ is the knob that tunes this balance.
  • $\lambda$ as a "Temperature" Parameter: The proposed annealing strategy for $\lambda$ is analogous to temperature in optimization algorithms like simulated annealing. Increasing $\lambda$ effectively "cools" the system, forcing it to adhere more strictly to the ODE model, which can help refine parameter estimates and avoid issues with overly flexible smooths.
  • Overcoming Limitations of Previous Methods: The authors demonstrate how their approach overcomes limitations of prior work, such as the need for known initial conditions or the use of overly simplistic bases that can introduce bias.
  • Robustness to Model Imperfection: The method shows promise in handling situations where the ODE model itself might not be perfectly accurate, by allowing for a controlled deviation from strict adherence to the model.

Key Results

  • The proposed method successfully estimates parameters and states for the nylon production system while ensuring that estimated concentrations remain non-negative.
  • The method demonstrates robustness to challenges like noisy data, unobserved states, and step changes in input variables.
  • The annealing strategy for $\lambda$ effectively refines parameter estimates and provides stable results.
  • Simulations show that the method's parameter estimates remain reasonably accurate even when the underlying ODE model is slightly misspecified.
  • Comparisons with other methods (e.g., NLS with fixed initial conditions, Bayesian approaches) show comparable or improved results, particularly in handling unobserved states and constraints.

Limitations and Future Work

  • Selection of $\lambda$: While an intuitive strategy for selecting $\lambda$ is proposed, the authors acknowledge that choosing the optimal $\lambda$ can still be an open problem, especially for complex systems.
  • Basis Function Choice: The performance of GS is influenced by the choice of basis functions. The paper notes that a denser basis might be needed for systems with very fast dynamics.
  • Model Parameterization: The paper briefly touches upon potential issues with parameterization, where certain parameters (like $\beta$ and $\gamma$ in the nylon model) can become highly correlated, making them difficult to uniquely identify. This suggests that careful model formulation is still important.
  • Future Work: The authors suggest extending the method to accommodate different error distributions beyond Gaussian and exploring its application to stochastic differential equation (SDE) models. Further investigation into optimal $\lambda$ selection and handling of more extreme model misspecifications are also implied.

Summary updated

Abstract

We introduce a method to estimate parameters and states from a differential equation model while enforcing interpretability constraints such as monotone or non‐negative states. We motivate the methodology using a real data chemical engineering example and show that a variety of restrictive constraints from earlier analyses do not address the problem of interpretability. Our proposed method estimates parameters using a smoothing‐based relaxation of the model to enforce interpretability of the observed and unobserved system states. Copyright © 2012 John Wiley & Sons, Ltd.
BibTeX
@article{campbell2012parameter,
  author = {David A. Campbell and Giles Hooker and Kimberley B. McAuley},
  title = {{Parameter estimation in differential equation models with constrained states}},
  journal = {Journal of Chemometrics},
  year = {2012},
  doi = {10.1002/cem.2416},
}
Figure