Bayesian Parameter Estimation for Nonlinear Dynamics Using Sensitivity Analysis

Yi Chou, Sriram SankaranarayananIJCAI 2019 · DOI

TL;DR

This paper introduces a novel Bayesian inference method for estimating parameters in nonlinear dynamical systems (like those described by ODEs). Instead of relying solely on computationally expensive sampling methods, it uses reachability analysis to derive bounds on the likelihood of parameters, offering a more robust and potentially faster approach, especially for systems with many parameters.

What Problem Does This Paper Address?

Many real-world systems in science and engineering, from chemical reactions to biological processes, are modeled using ordinary differential equations (ODEs). However, the parameters within these ODEs are often not directly measurable and significantly influence the system's behavior. Estimating these parameters from noisy observational data is crucial. Traditional methods like Markov Chain Monte Carlo (MCMC) and Sequential Monte Carlo (SMC) are widely used but can be computationally intensive, especially as the number of parameters increases (the "curse of dimensionality"). This paper aims to develop a Bayesian inference technique that can provide accurate parameter estimates for nonlinear ODEs while mitigating the computational burden associated with high-dimensional parameter spaces.

Methods

The core of this paper's approach lies in combining reachability analysis with sensitivity analysis within a Bayesian framework.

  • Bayesian Inference: The standard Bayesian approach is used, where a prior distribution over the parameters is updated with observational data to obtain a posterior distribution. The challenge is computing this posterior, especially the likelihood term $P(Y|p)$.
  • Reachability Analysis: Instead of directly computing the likelihood for every parameter value, the authors use set-valued reachability analysis. This technique computes guaranteed bounds on the system's state and output trajectories for a given set of parameter values. This allows them to derive interval bounds on the likelihood function for subsets (cells) of the parameter space.
  • Sensitivity Analysis: To make reachability analysis more efficient, especially for large parameter spaces, the paper employs a linearized sensitivity analysis. This involves computing a Jacobian matrix to approximate how outputs change with small perturbations in parameters. While this approximation is "unsound" (meaning it doesn't provide strict guarantees like full reachability analysis), it's significantly faster. The authors argue that for cells with low likelihood, a coarse approximation is sufficient, and for cells with high likelihood, the adaptive refinement process makes the linearization precise enough.
  • Adaptive Partitioning: The parameter space is initially partitioned into cells. The likelihood bounds for each cell are computed. Cells with high uncertainty (large difference between upper and lower likelihood bounds) or high probability are then adaptively refined (subdivided) to improve the accuracy of the posterior approximation. This focuses computational effort on the most relevant parts of the parameter space.

Datasets

The paper evaluates its approach on several benchmarks:

  • Simple Ball Example: A basic physics example with one parameter ($g$).
  • Fitzhugh-Nagumo Model: A nonlinear model for neuronal dynamics with two parameters ($\theta_1, \theta_2$).
  • Rossler Benchmark: A chaotic system with two parameters.
  • Laub-Loomis Models (model-1 to model-7): A series of models with varying numbers of parameters (3 to 6) and state variables.
  • GENETIC Benchmark: A model with two parameters.
  • Dalla-Man Model: A two-parameter model known for strong parameter correlation.
  • P53 Benchmark: A model with two parameters.
  • ANALOG Benchmark: A simple model with one parameter.

No new datasets were introduced; existing benchmarks were used to compare against established methods.

Key Insights and Ideas

  • Bridging Reachability and Bayesian Inference: The paper ingeniously connects rigorous set-valued reachability analysis with the probabilistic world of Bayesian inference. This allows for guaranteed bounds on the posterior, which is a significant advantage over purely sampling-based methods that only provide probabilistic estimates.
  • Efficiency through Approximation: The use of linearized sensitivity analysis, while not strictly sound, is a pragmatic choice that dramatically speeds up the computation for large parameter spaces. The adaptive refinement strategy ensures that this approximation is applied judiciously.
  • Imprecise Probability for Posterior: The output is an "imprecise probability distribution" represented by upper and lower bounds on the posterior probability of parameter subsets. This is a powerful way to represent uncertainty when exact computation is intractable.
  • Curse of Dimensionality Mitigation: The adaptive partitioning and sensitivity analysis are specifically designed to combat the curse of dimensionality, showing promise for systems with up to 6 parameters.

Key Results

  • The proposed approach yields an accurate description of the posterior distribution, comparable to MCMC and SMC methods.
  • The computational time for their approach is competitive with MCMC, especially when considering the number of samples required by MCMC to achieve comparable variance.
  • The method demonstrates effectiveness for systems with up to 6 parameters.
  • Beyond 6 parameters, the approach suffers from the curse of dimensionality, indicating a limitation.
  • The paper provides quantitative comparisons in Table 1, showing computation times and estimated posterior expectations ($E(p)$) for various benchmarks, demonstrating good agreement with MCMC/SMC results.

Limitations and Future Work

  • Curse of Dimensionality: The authors acknowledge that their approach struggles with parameter spaces beyond 6 dimensions.
  • Unsound Approximation: The use of linearized sensitivity analysis, while efficient, is not strictly sound and could lead to inaccuracies in certain scenarios if not managed by the refinement process.
  • Future Work: The authors suggest exploring extensions to probabilistic programming languages for hybrid cyber-physical systems. They also propose investigating methods to factor the joint distribution over many parameters into simpler distributions over subsets, potentially by analyzing parameter correlations, to further improve scalability.

Summary updated

BibTeX
@inproceedings{chou2019bayesian,
  author = {Yi Chou and Sriram Sankaranarayanan},
  title = {{Bayesian Parameter Estimation for Nonlinear Dynamics Using Sensitivity Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/791},
}
Figure