MFBO-SSM: Multi-Fidelity Bayesian Optimization for Fast Inference in State-Space Models

Mahdi Imani, Seyede Fatemeh Ghoreishi, Douglas L. Allaire, Ulisses Braga-NetoAAAI 2019 · OpenReview

Abstract

Nonlinear state-space models are ubiquitous in modeling real-world dynamical systems. Sequential Monte Carlo (SMC) techniques, also known as particle methods, are a well-known class of parameter estimation methods for this general class of state-space models. Existing SMC-based techniques rely on excessive sampling of the parameter space, which makes their computation intractable for large systems or tall data sets. Bayesian optimization techniques have been used for fast inference in state-space models with intractable likelihoods. These techniques aim to find the maximum of the likelihood function by sequential sampling of the parameter space through a single SMC approximator. Various SMC approximators with different fidelities and computational costs are often available for sample-based likelihood approximation. In this paper, we propose a multi-fidelity Bayesian optimization algorithm for the inference of general nonlinear state-space models (MFBO-SSM), which enables simultaneous sequential selection of parameters and approximators. The accuracy and speed of the algorithm are demonstrated by numerical experiments using synthetic gene expression data from a gene regulatory network model and real data from the VIX stock price index.

Introduction

Parameter estimation in nonlinear dynamical systems is a classic "expensive" problem. Whether modeling the volatility of the stock market or the complex interactions within a gene regulatory network, researchers frequently rely on State-Space Models (SSMs). The standard for inference in these models is Sequential Monte Carlo (SMC)—better known as particle filtering.

However, particle filters are computationally demanding. To obtain an accurate estimate of the likelihood, a high number of particles is required. When running optimization or sampling routines that necessitate thousands of likelihood evaluations, the computational clock starts ticking fast. In nonlinear SSMs, the likelihood function is usually intractable and must be approximated. This approximation is directly tied to the number of particles ($N$): high $N$ is accurate but slow, while low $N$ is fast but noisy. Standard Bayesian Optimization (BO) techniques typically use a fixed, high $N$ to ensure accuracy, which makes the optimization process sluggish. MFBO-SSM addresses this bottleneck by optimizing not just the parameters, but the "fidelity" of the simulations themselves.

Methods: Multi-Fidelity Bayesian Optimization

The core innovation of MFBO-SSM lies in treating the particle count $N$ as a controllable resource. The algorithm intelligently switches between fast, "low-fidelity" particle filters for exploration and expensive, "high-fidelity" ones for pinpointing the optimal solution.

Comparison of log-likelihood approximations using low-fidelity (left) and high-fidelity (right) particle filters. Low-fidelity filters are computationally cheap but noisy, whereas high-fidelity filters provide a more accurate estimate of the exact log-likelihood.

The framework utilizes three key components:

  1. Gaussian Process (GP) Surrogate: Instead of evaluating the expensive likelihood function directly, the algorithm builds a GP model of the log-likelihood surface that accounts for the noise inherent in particle filtering.
  2. Multi-Fidelity Modeling: The GP treats the number of particles $N$ as a controllable fidelity level. Each level has a known computational cost and a corresponding noise variance.
  3. Knowledge Gradient (KG) Policy: This policy decides both where to sample and how many particles to use. It selects the parameter-fidelity pair that maximizes the expected increase in the inference function per unit cost.

Experimental Evaluation

The authors evaluated MFBO-SSM on three distinct datasets:

  • Synthetic Stochastic Volatility Model: A standard benchmark in financial econometrics used to test handling of long time-series data.
  • Real Financial Data: Daily returns of the VIX stock price index (2005–2018), testing robustness on noisy, real-world signals.
  • Cell-Cycle Gene Regulatory Network: A complex 14-gene Boolean network model representing a high-dimensional discrete state-space problem common in systems biology.

Results and Discussion

Across all datasets, MFBO-SSM proved to be both faster and more accurate than traditional methods.

Maximum Log-Likelihood vs. Running Time. MFBO-SSM achieves a higher maximum log-likelihood significantly faster than standard BO, EM, and other baseline inference methods.

Key results include:

  • Speedup: In the gene regulatory network example, MFBO-SSM was 2.5x to 5x faster than standard BO or Particle Marginal Metropolis-Hastings (PMMH) while achieving lower Mean Squared Error (MSE).
  • Optimization Efficiency: On VIX data, MFBO-SSM found a higher maximum log-likelihood significantly faster than traditional Expectation-Maximization (EM) or direct Maximum Likelihood (ML) approaches.
  • Dynamic Resource Allocation: Visualizations show the algorithm predominantly picking low-fidelity samples for initial exploration, only switching to high-fidelity samples at the end to refine the peak.

Conclusion

MFBO-SSM demonstrates that treating simulation fidelity as a budget item can significantly accelerate parameter estimation. By formalizing the cost-benefit trade-off of particle counts, the algorithm automates a strategy that researchers often attempt manually. Future work may focus on improving performance in extremely high-dimensional spaces and automating noise variance estimation across diverse model types.

Summary updated

Abstract

Nonlinear state-space models are ubiquitous in modeling real-world dynamical systems. Sequential Monte Carlo (SMC) techniques, also known as particle methods, are a well-known class of parameter estimation methods for this general class of state-space models. Existing SMC-based techniques rely on excessive sampling of the parameter space, which makes their computation intractable for large systems or tall data sets. Bayesian optimization techniques have been used for fast inference in state-space models with intractable likelihoods. These techniques aim to find the maximum of the likelihood function by sequential sampling of the parameter space through a single SMC approximator. Various SMC approximators with different fidelities and computational costs are often available for sample-based likelihood approximation. In this paper, we propose a multi-fidelity Bayesian optimization algorithm for the inference of general nonlinear state-space models (MFBO-SSM), which enables simultaneous sequential selection of parameters and approximators. The accuracy and speed of the algorithm are demonstrated by numerical experiments using synthetic gene expression data from a gene regulatory network model and real data from the VIX stock price index.
BibTeX
@inproceedings{imani2019mfbo,
  author = {Mahdi Imani and Seyede Fatemeh Ghoreishi and Douglas L. Allaire and Ulisses Braga-Neto},
  title = {{MFBO-SSM: Multi-Fidelity Bayesian Optimization for Fast Inference in State-Space Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=BtV4nlaTqo},
}
Figure