[Paper Review] Guiding the Sequential Experiments in Autonomous Experimentation Platforms through EI-based Bayesian Optimization and Bayesian Model Averaging

Ahmed Shoyeb Raihan, Imtiaz Ahmed|arXiv (Cornell University)|Feb 26, 2023

Fault Detection and Control Systems4 citations

TL;DR

This paper proposes a Bayesian optimization framework enhanced with Bayesian Model Averaging (BMA) to improve sequential experiment selection in autonomous experimentation platforms (AEPs). By integrating Expected Improvement (EI)-based acquisition with model averaging across multiple predictive models, the method reduces uncertainty and improves prediction accuracy on a real steel fatigue strength dataset, outperforming traditional single-model BO in RMSE.

ABSTRACT

Autonomous Experimentation Platforms (AEPs) are advanced manufacturing platforms that, under intelligent control, can sequentially search the material design space (MDS) and identify parameters with the desired properties. At the heart of the intelligent control of these AEPs is the policy guiding the sequential experiments, which is to choose the location to carry out the next experiment. In such cases, a balance between exploitation and exploration must be achieved. A Bayesian Optimization (BO) framework with Expected Improvement based (EI-based) acquisition function can effectively search the MDS and guide where to conduct the next experiments so that the underlying relationship can be identified with a smaller number of experiments. The traditional BO framework tries to optimize a black box objective function in a sequential manner by relying on a single model. However, this single-model approach does not account for model uncertainty. Bayesian Model Averaging (BMA) addresses this issue by working with multiple models and thus considering the uncertainty in the models. In this work, we first apply the conventional BO algorithm with the most popular EI-based experiment policy in a real-life fatigue dataset for steel to predict the fatigue strength of steel. Afterward, we apply BMA to the same dataset by working with a set of predictive models and compare the performance of BMA with the traditional BO algorithm, which relies on a single model for approximation. We compare the results in terms of RMSE and find that BMA performs better than EI-based BO in the prediction task by considering the model uncertainty in its framework.

Motivation & Objective

• To improve sequential experiment selection in autonomous experimentation platforms (AEPs) by reducing model uncertainty in Bayesian optimization.
• To evaluate the performance of Bayesian Model Averaging (BMA) against traditional Expected Improvement (EI)-based Bayesian Optimization (BO) in real-world material design.
• To assess whether incorporating model uncertainty through BMA leads to better prediction accuracy in material property optimization.
• To demonstrate the effectiveness of BMA-BO in minimizing experimental iterations while improving prediction reliability on a real fatigue strength dataset.

Proposed method

• The study applies standard Expected Improvement (EI)-based Bayesian Optimization (BO) using a single Gaussian Process (GP) model to guide sequential experiments on a real steel fatigue dataset.
• It extends the BO framework by integrating Bayesian Model Averaging (BMA), which combines predictions from multiple GP models to account for model uncertainty.
• BMA computes a weighted average of predictions from multiple models, where weights are based on model posterior probabilities, improving robustness to model misspecification.
• The acquisition function in the BMA-BO framework is derived from the expected improvement under the predictive distribution of the averaged model.
• The method sequentially selects the next experiment location by maximizing the expected improvement over the current best observation, using the BMA-predicted mean and variance.
• Performance is evaluated using RMSE on the test set, comparing BMA-BO against standard EI-based BO with a single GP model.

Experimental results

Research questions

• RQ1 Does integrating Bayesian Model Averaging (BMA) into Expected Improvement-based Bayesian Optimization improve prediction accuracy in autonomous experimentation?
• RQ2 How does accounting for model uncertainty via BMA affect the convergence and robustness of sequential experiment selection in material design?
• RQ3 Can BMA-BO reduce the number of required experiments while maintaining or improving prediction performance on real material property datasets?
• RQ4 How does BMA-BO compare to single-model BO in terms of RMSE on a real-world steel fatigue strength dataset?

Key findings

• Bayesian Model Averaging (BMA) significantly improves prediction accuracy compared to traditional EI-based Bayesian Optimization with a single model.
• The BMA-BO framework achieves a lower Root Mean Square Error (RMSE) on the test set than the single-model BO approach, demonstrating improved generalization.
• By incorporating uncertainty across multiple models, BMA-BO reduces the risk of poor predictions due to model misspecification.
• The integration of BMA into the acquisition function leads to more robust and reliable experiment selection in sequential material discovery.