PaperlayerView exact revision on arXiv

Source-linked AI summary

Simulation-based optimal Bayesian experimental design for nonlinear systems

Xun Huan, Youssef M. Marzouk

arXiv:1108.4146v3stat.MLstat.COstat.ME

TL;DR

The paper addresses how to choose informative experiments for parameter inference when nonlinear simulation-based models make design evaluation expensive. It develops a Bayesian information-theoretic framework with polynomial chaos surrogates, two-stage Monte Carlo estimation, and stochastic optimization, and demonstrates informative designs with substantial computational savings.

  • Problem

    The problem is selecting experimental conditions that maximize data value for inference when experiments are time-consuming, expensive, or delicate.

  • Method

    The method combines Bayesian parameter inference, expected information gain, polynomial chaos surrogates, two-stage Monte Carlo estimation, and stochastic optimization.

  • Results

    More than three orders of magnitude in computational savings were obtained with surrogates over stochastic optimization using the full model in the tested examples.

  • Takeaways & Limitations

    Optimal design conditions were highly informative about targeted parameters and more informative than experiments selected with simple heuristics.

  • Takeaways & Limitations

    Sequential experimental design over many experiments is outside the paper’s scope, and the nested Monte Carlo estimator is biased.

Abstract

from arXiv · show

The optimal selection of experimental conditions is essential to maximizing the value of data for inference and prediction, particularly in situations where experiments are time-consuming and expensive to conduct. We propose a general mathematical framework and an algorithmic approach for optimal experimental design with nonlinear simulation-based models; in particular, we focus on finding sets of experiments that provide the most information about targeted sets of parameters. Our framework employs a Bayesian statistical setting, which provides a foundation for inference from noisy, indirect, and incomplete data, and a natural mechanism for incorporating heterogeneous sources of information. An objective function is constructed from information theoretic measures, reflecting expected information gain from proposed combinations of experiments. Polynomial chaos approximations and a two-stage Monte Carlo sampling method are used to evaluate the expected information gain. Stochastic approximation algorithms are then used to make optimization feasible in computationally intensive and high-dimensional settings. These algorithms are demonstrated on model problems and on nonlinear parameter estimation problems arising in detailed combustion kinetics.

1. Introduction

The paper targets optimal experimental design for nonlinear, computationally intensive systems, where experiments are costly and conventional criteria require simplifying assumptions. It combines Bayesian information-theoretic objectives with approximation and stochastic optimization strategies.

  • Motivation and gap: Nonlinear optimal-design criteria are difficult to evaluate exactly, motivating assumptions such as model linearization and Gaussian posterior approximations.These assumptions can change the form of the design objective.
  • Contribution: The paper addresses computationally intensive nonlinear systems with flexible strategies that use a full information-theoretic formalism and few limiting assumptions.The stated goal is to obtain optimal experimental designs efficiently.
  • Approach: The framework uses Bayesian inference, expected Shannon information gain, polynomial chaos surrogates, and stochastic optimization to select informative experimental conditions.The approach is designed for high-dimensional settings and can plan experiments without discretizing the design space.
  • Framework: The framework organizes design criteria, objective estimation, stochastic optimization, surrogate construction, inference, and experimentation within a design–experimentation–model improvement cycle.Figure 1 presents these components in a flowchart.

2. Experimental design formulation

The formulation defines Bayesian experimental design for parameter inference using expected information gain, then makes its evaluation and optimization tractable for nonlinear models. It also treats multiple experiments and identifies computational and algorithmic limitations.

  • 2.1. Experimental goals: The experimental goal is inference of a finite number of model parameters, while utility functions may also include penalties for effort, cost, or resource constraints.The framework can be generalized to other experimental goals.
  • 2.2. Design criterion and expected utility: Bayesian design supports inference from noisy, indirect, and incomplete data while incorporating physical constraints and heterogeneous information sources.Unknown parameters and data are represented probabilistically under the Bayesian paradigm.
  • 2.2. Design criterion and expected utility: The expected utility averages a utility function over uncertain parameters and possible observations for each experimental design.The utility reflects the usefulness of an experiment at specified conditions and outcomes.
  • 2.2. Design criterion and expected utility: The paper chooses posterior-to-prior KL divergence as utility, so expected utility measures average information gain about the parameters and equals their mutual information with the data.Designs are selected by maximizing this expected utility over the design space.
  • 2.2. Design criterion and expected utility: Simultaneously planned experiments are optimized jointly because repeating the single-experiment optimum does not generally maximize total expected information gain.Their data are incorporated together in an augmented likelihood.
  • 2.2. Design criterion and expected utility: Sequential design can use observed results to update the prior for later experiments, but rigorous optimization over many experiments through dynamic programming is outside the paper’s scope.The described greedy strategy is not necessarily optimal over a long horizon.
  • 2.3. Numerical evaluation of the expected utility: Expected utility generally lacks a closed form, and under a design-independent joint entropy assumption, maximizing data entropy is equivalent to maximizing expected information gain.The latter simplification applies only in the stated special case.
  • 2.4. Stochastic optimization: The nested Monte Carlo estimator is biased, with inner samples controlling bias and outer samples controlling variance; stochastic optimization is needed because noisy objective evaluations make grid and deterministic searches expensive.SPSA uses two evaluations per step but can stagnate under high noise, whereas NMNS is less noise-sensitive.

3. Polynomial chaos surrogate

The paper replaces expensive forward-model evaluations with polynomial chaos surrogates that cover parameter and design spaces, then computes their coefficients non-intrusively using quadrature. Dimension-adaptive sparse quadrature reduces the integration burden in moderate-dimensional settings.

  • Motivation: Expensive forward-model evaluations make expected information-gain optimization impractical, motivating cheaper surrogates accurate across the prior support and design space.The surrogate replaces repeated evaluations of G while retaining dependence on uncertain parameters and design conditions.
  • Polynomial chaos construction: Polynomial chaos expansions exploit regularity in model outputs and are constructed jointly over uncertain parameters and design conditions.A single expansion for each component of G avoids constructing separate expansions at every design value considered during optimization.
  • Polynomial chaos construction: Total-order truncation limits stochastic dimension and polynomial order, but the number of expansion terms grows rapidly as these parameters increase.The truncation order reflects nonlinearity, while stochastic dimension reflects the degrees of freedom needed to capture system stochasticity.
  • Pseudospectral projection: Non-intrusive spectral projection computes polynomial chaos coefficients using numerical quadrature because the forward model appears in the coefficient integrals.The coefficient denominators are available analytically, whereas the numerators require forward-model evaluations.
  • Sparse quadrature: Dimension-adaptive sparse quadrature adaptively tensorizes one-dimensional rules and has weak dependence on dimension, making it suitable for moderate-size problems.The method uses active and old multi-index sets, while Clenshaw-Curtis quadrature is selected for its accuracy, nestedness, and construction simplicity.

4. Bayesian parameter inference

Bayesian parameter inference uses posterior sampling when nonlinear models make direct posterior evaluation or grid computation impractical. MCMC provides posterior samples, while surrogate forward models can reduce the cost of the many model evaluations required.

  • Inference objective: The inference goal is to characterize a narrow posterior distribution for the model parameters after collecting data from an optimal experiment.A narrow posterior indicates that the data constrain the plausible parameter range.
  • Posterior computation: Grid-based posterior computation becomes impractical in high dimensions, and arbitrary nonlinear posterior forms generally prevent direct Monte Carlo sampling.The paper therefore turns to Markov chain Monte Carlo methods.
  • Posterior computation: MCMC constructs a Markov chain whose stationary distribution is the posterior using pointwise evaluations of the unnormalized posterior density.The resulting samples are correlated, so their effective sample size is smaller than the number of MCMC steps.
  • Posterior summaries: Posterior samples support estimates such as the MMSE estimator and Bayes risk, represented respectively by the posterior mean and variance.Both quantities can be estimated from MCMC samples.
  • Algorithmic efficiency: DRAM combines delayed rejection and adaptive Metropolis with Metropolis-Hastings to balance simplicity and efficiency in this setting.MCMC may require tens of thousands or millions of samples, making polynomial chaos surrogates valuable for reducing repeated forward-model costs.

5. Application: nonlinear model

A nonlinear scalar model illustrates single- and two-experiment Bayesian design under a uniform prior. The best single designs occur near 0.2 and 1.0, while the best pair combines these conditions over the full prior.

  • Single-experiment design: The study uses expected utility and its estimate to maximize expected information about θ from a single measurement.The inexpensive algebraic example demonstrates expected information-gain estimation and the role of prior information.
  • Model setup: The illustrative model has one scalar observable, one uncertain parameter, one design variable, additive Gaussian noise, and a uniform prior on θ over [0, 1].The design space is d ∈[0, 1], and the noise variance is 10^-4.
  • Single-experiment design: Local utility maxima occur at d = 0.2 and d = 1.0 because the model signal dominates measurement noise near either condition.At other design values, the observation is dominated by noise and is less useful for inferring θ.
  • Two-experiment design: For two fixed experiments under the full prior, the optimal pair is (0.2, 1.0) or (1.0, 0.2), rather than repeating one single-experiment optimum.The two designs have greater slopes over complementary portions of the prior range.
  • Two-experiment design: Restricted priors favor repeating the corresponding single-experiment optimum, whereas the full prior favors combining different conditions.Design 0.2 is optimal over [0, θe], and design 1.0 is optimal over [θe, 1].

6. Application: combustion kinetics

The combustion-kinetics application uses shock-tube ignition simulations to design experiments for inferring selected kinetic parameters. Polynomial-chaos surrogates and stochastic optimization identify informative designs while substantially reducing computational cost, although surrogate accuracy requirements differ between design and inference.

  • Experimental setup: Shock-tube ignition experiments provide indirect information about elementary chemical kinetic parameters through ignition delays and other observables.The modeled hydrogen–oxygen system is governed by stiff, nonlinear ODEs, and the demonstration targets parameters A1 and Ea,3.
  • Polynomial chaos construction: The peak enthalpy release-rate polynomial-chaos expansion exhibits aliasing error at low nquad and exponential convergence once truncation error dominates.The recommended operating region is near the contour-plot knees, where increasing nquad yields little additional accuracy.
  • Single-experiment design: The full ODE model and polynomial-chaos surrogate identify the same optimal single-experiment design at approximately (T0*, φ*) = (975, 0.5).Their expected-utility contours are similar, and the surrogate’s posterior contours match the full model closely.
  • Single-experiment design: The posterior is tightest at design A, making it the most informative of the three tested conditions, while posterior modes differ because of surrogate modeling error and likelihood noise.The full-model posterior modes also need not equal the artificial-data parameters when the likelihood includes noise.
  • Observable selection: Characteristic-time observables are more informative than peak-value observables, and changing the observable set can substantially alter the optimal design parameters.Expected utilities from partitioned observable sets do not simply add to the full-observable utility.
  • Two-experiment optimization: Both SPSA and NMNS place two experiments near T0 = 975 K, but SPSA produces tighter clustering with more outliers and NMNS reaches its plateau faster at lower nout.Increasing nout generally tightens final-design clustering but requires more model evaluations per iteration.
  • Computational cost and surrogate accuracy: A full-model optimization requiring 5000 × 10^4 = 5 × 10^7 evaluations can be replaced by constructing a surrogate with 10^4 full ODE evaluations.A low-quality surrogate may still locate good designs, but p = 6, nquad = 10^4 produces substantially less accurate inference posteriors.

7. Conclusions

The paper presents a Bayesian framework for optimal experimental design with nonlinear, computationally intensive models, combining information-theoretic objectives, Monte Carlo estimation, surrogates, and stochastic optimization. Applications show informative model-based designs and major computational savings, while future work highlights uncertainty in design conditions and structural model inadequacy.

  • Framework: The framework links nonlinear, computationally intensive experiment models to rigorous information-theoretic criteria without requiring essentially simplifying assumptions.Its overall workflow places optimal design within a design–experimentation–model-improvement cycle.
  • Computational approach: Expected information gain for Bayesian parameter inference is estimated with a two-stage Monte Carlo method and optimized using stochastic optimization algorithms.Flexible polynomial-chaos surrogates substantially accelerate otherwise prohibitive objective evaluations.
  • Demonstrations: The method is demonstrated on a nonlinear algebraic model and shock-tube ignition experiments governed by stiff nonlinear ODEs, including single- and multiple-experiment designs.The combustion example is challenging because observables can depend sharply on kinetic parameters and design conditions.
  • Main findings: Surrogate-based optimal designs are highly informative for targeted parameters and more informative than designs obtained with simple heuristics.Across the tested examples, surrogates provide computational savings exceeding three orders of magnitude relative to full-model stochastic optimization.
  • Main findings: Surrogates need not be extremely accurate to reveal correct optimal design points, but they require greater accuracy for parameter inference.The paper identifies surrogate accuracy as a distinct requirement for design and inference tasks.
  • Future work: Future work should address uncertainly imposed design conditions, structural model inadequacy, more efficient adaptive polynomial-chaos construction, and sequential Bayesian design.The authors state that structural uncertainty requires substantially more investigation.

Appendix A. Expected information gain from two experiments

The appendix analyzes how expected information gain behaves when two experiments are performed together. It shows that combined information is bounded by the sum of individual gains, with equality only when the experiments provide no overlapping information.

  • Combined experiments: Expected information gain from two experiments is not generally equal to the sum of their individual expected information gains.The appendix derives this result for fixed conditions d1 and d2 with corresponding data y1 and y2.
  • Derivation: Conditional independence of the experiment outputs given designs and parameters enables the Bayesian factorization used in the derivation.The derivation also marginalizes over θ to express the result using differential entropy.
  • Result: U([d1, d2]) ≤ U(d1) + U(d2), with equality only when the data have zero conditional mutual information and therefore no overlapping information.The inequality means simultaneous experiments cannot provide more total expected information than the sum of their separate gains.

Appendix B. Bias in the estimator of expected information gain

The expected information gain estimator is biased by finite inner sampling and reused samples, but numerical tests find the bias small and sample reuse computationally worthwhile.

  • The estimator has two bias sources: finite n_in and reuse of prior samples between outer and inner Monte Carlo loops.Sample reuse reduces distinct evaluations of the forward model G(θ, d).
  • At n_in = 10^6, close agreement between reuse and resample estimates suggests extremely small bias.The resample estimator at n_in = 10^6 is treated as the reference for assessing bias.
  • For n_in = 10^4 or 10^5, the bias falls by another order of magnitude or more.
  • Sample reuse provides substantial computational efficiency through fewer evaluations of G, enabling larger n_in for the same computational effort.
  • Small, approximately design-stationary bias is expected to have a small overall effect on stochastic optimization results.The authors note that optimal designs should not change dramatically when bias-induced shifts preserve relative expected utilities among designs.

Appendix C. Governing equations for homogeneous combustion

The appendix formulates homogeneous, constant-pressure combustion as coupled ODEs for species and energy conservation, with reaction rates governed by modified Arrhenius kinetics.

  • The combustion system is spatially homogeneous and constant-pressure, with convective and diffusive transport neglected.The resulting zero-dimensional model uses coupled ordinary differential equations for species and energy conservation.
  • The system state consists of species mass fractions Y_1, ..., Y_ns and temperature T.
  • Species production rates are defined from elementary reaction rates, stoichiometric coefficients, and species molar concentrations.Molar concentrations are obtained from species mass fractions.
  • Forward and reverse reaction-rate constants use a modified Arrhenius form with kinetic parameters A_m, b_m, and E_a,m.A_m is the pre-exponential factor, b_m the temperature-dependence exponent, and E_a,m the activation energy.
  • Initial mass fractions are specified using the dimensionless equivalence ratio φ, while the model uses species molar fractions X_j as state variables.The mixture is closed with a perfect-gas equation of state at assumed constant pressure.
Loading 1108.4146v3…