Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Bayesian learning”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 253 records · Page 14
All sourcesNASA NTRSDOE OSTI
DOE OSTI2020

Interpreting High-resolution Spectroscopy of Exoplanets using Cross-correlations and Supervised Machine Learning

We present a new method for performing atmospheric retrieval on ground-based, high-resolution data of exoplanets. Our method combines cross-correlation functions with a random forest, a supervised machine-learning technique, to overcome challenges associated with high-resolution data. A series of cross-correlation functions are concatenated to give a “CCF-sequence” for each model atmosphere, which reduces the dimensionality by a factor of ∼100. The random forest, trained on our grid of ∼65,000 models, provides a likelihood-free method of retrieval. The precomputed grid spans 31 values of both temperature and metallicity, and incorporates a realistic noise model. We apply our method to HARPS-N observations of the ultra-hot Jupiter KELT-9b and obtain a metallicity consistent with solar (logM = − 0.2 ± 0.2). Our retrieved transit chord temperature (T=6000{sub −200}{sup +0}K) is unreliable as strong ion lines lie outside of the extent of the training set, which we interpret as being indicative of missing physics in our atmospheric model. We compare our method to traditional nested sampling, as well as other machine-learning techniques, such as Bayesian neural networks. We demonstrate that the likelihood-free aspect of the random forest makes it more robust than nested sampling to different error distributions, and that the Bayesian neural network we tested is unable to reproduce complex posteriors. We also address the claim in Cobb et al. 2019 that our random forest retrieval technique can be overconfident but incorrect. We show that this is an artifact of the training set, rather than of the machine-learning method, and that the posteriors agree with those obtained using nested sampling.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Active Learning Guided Computational Discovery of Plant-Based Redoxmers for Organic Nonaqueous Redox Flow Batteries

Organic nonaqueous redox flow batteries (ONRFBs) are promising energy storage devices due to their scalability and reliance on sourceable materials. However, finding suitable redox-active organic molecules (redoxmers) for these batteries remains a challenge. Using plant-based compounds as precursors for these redoxmers can decrease their costs and environmental toxicity. In this computational study, flavonoid molecules have been examined as potential redoxmers for ONRFBs. Flavone and isoflavone derivatives were selected as catholyte (positive charge carrier) and anolyte (negative charge carrier) molecules, respectively. To drive their redox potentials to the opposite extremes, in silico derivatization was performed using a novel algorithm to generate a library of > 40000 candidate molecules that penalizes overly complex structures. A multiobjective Bayesian optimization based active learning algorithm was then used to identify best redoxmer candidates in these search spaces. Furthermore, our study provides methodologies for molecular design and optimization of natural scaffolds and highlights the need of incorporating expert chemistry awareness of the natural products and the basic rules of synthetic chemistry in machine learning.

25 ENERGY STORAGE↗
DOE OSTI2026

A multi-backend autotuning study of feature selection on GPUs

Abstract Feature selection is an important step in machine learning that can benefit from GPU acceleration. As the number of GPU vendors increases, it is imperative to adapt algorithms such as the minimum Redundancy Maximum Relevance (mRMR) feature selection method to different backends that support several GPU architectures. This work presents a multi-backend implementation of mRMR across CUDA, HIP, and SYCL, and studies its performance when combined with Bayesian optimization and transfer learning to automatically tune execution parameters for different platforms and datasets. Our experimental results show that when tuned, CUDA and HIP achieve comparable performance on NVIDIA architectures, while SYCL exhibits a moderate performance gap. Overall, this work highlights the impact of backend choice and autotuning on GPU-accelerated feature selection and provides insights into deploying mRMR across heterogeneous environments.

Beceiro, Bieito (ORCID:0000000333014890)↗
DOE OSTI2018

Active learning for accelerated design of layered materials

Hetero-structures made from vertically stacked monolayers of transition metal dichalcogenides hold great potential for optoelectronic and thermoelectric devices. Discovery of the optimal layered material for specific applications necessitates the estimation of key material properties, such as electronic band structure and thermal transport coefficients. However, screening of material properties via brute force ab initio calculations of the entire material structure space exceeds the limits of current computing resources. Moreover, the functional dependence of material properties on the structures is often complicated, making simplistic statistical procedures for prediction difficult to employ without large amounts of data collection. Here, we present a Gaussian process regression model, which predicts material properties of an input hetero-structure, as well as an active learning model based on Bayesian optimization, which can efficiently discover the optimal hetero-structure using a minimal number of ab initio calculations. The electronic band gap, conduction/valence band dispersions, and thermoelectric performance are used as representative material properties for prediction and optimization. The Materials Project platform is used for electronic structure computation, while the BoltzTraP code is used to compute thermoelectric properties. Bayesian optimization is shown to significantly reduce the computational cost of discovering the optimal structure when compared with finding an optimal structure by building a regression model to predict material properties. The models can be used for predictions with respect to any material property and our software, including data preparation code based on the Python Materials Genomics (PyMatGen) library as well as python-based machine learning code, is available open source.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Atikokan Digital Twin: Machine learning in a biomass energy system

The Atikokan Generating Station, operated by Ontario Power Generation, has a 200 MW, biomass-fired tower boiler that operates on a dispatch schedule with a five-minute cycle. The boiler is generally operated in the range of 40–100 MW using two of five burner levels. In order to optimize boiler performance, we propose the implementation of a unique digital twin. Our digital twin abstraction couples Bayesian inference from science-based models and from observations (machine learning) with decision theory to predict operating-variable set points that optimize the physical asset (the boiler) in the presence of uncertainty (artificial intelligence). We focus this paper on the continuous Bayesian machine learning part of the Atikokan Digital Twin; we discuss decision theory in a companion paper. We identify and learn about 12 operational, model, and measured-output parameters and their uncertainties from high-fidelity, science-based simulations of the Atikokan boiler and from the observed measurements at the power plant. Since the goal of the Atikokan Digital Twin is to implement it online in real time, we require fast function evaluations for the quantities of interest extracted from the simulations in the Bayesian analysis. We use Gaussian process regression/interpolation to create accurate, robust surrogate models. We define the Bayesian priors and likelihood function and solve for the posterior distributions of the 12 parameters. Here we then propagate these distributions (i.e., parameters with uncertainty) into the predicted distributions of 790 quantities of interest to learn about the relative importance of various sources of error including experimental, model, and operating-parameter errors.

09 BIOMASS FUELS↗
DOE OSTI2021

Data-Driven Strategies for Accelerated Materials Design

The ongoing revolution of the natural sciences by the advent of machine learning and artificial intelligence sparked significant interest in the material science community in recent years. The intrinsically high dimensionality of the space of realizable materials makes traditional approaches ineffective for large-scale explorations. Modern data science and machine learning tools developed for increasingly complicated problems are an attractive alternative. An imminent climate catastrophe calls for a clean energy transformation by overhauling current technologies within only several years of possible action available. Tackling this crisis requires the development of new materials at an unprecedented pace and scale. For example, organic photovoltaics have the potential to replace existing silicon-based materials to a large extent and open up new fields of application. In recent years, organic light-emitting diodes have emerged as state-of-the-art technology for digital screens and portable devices and are enabling new applications with flexible displays. Reticular frameworks allow the atom-precise synthesis of nanomaterials and promise to revolutionize the field by the potential to realize multifunctional nanoparticles with applications from gas storage, gas separation, and electrochemical energy storage to nanomedicine. In the recent decade, significant advances in all these fields have been facilitated by the comprehensive application of simulation and machine learning for property prediction, property optimization, and chemical space exploration enabled by considerable advances in computing power and algorithmic efficiency. In this Account, we review the most recent contributions of our group in this thriving field of machine learning for material science. We start with a summary of the most important material classes our group has been involved in, focusing on small molecules as organic electronic materials and crystalline materials. Specifically, we highlight the data-driven approaches we employed to speed up discovery and derive material design strategies. Subsequently, our focus lies on the data-driven methodologies our group has developed and employed, elaborating on high-throughput virtual screening, inverse molecular design, Bayesian optimization, and supervised learning. We discuss the general ideas, their working principles, and their use cases with examples of successful implementations in data-driven material discovery and design efforts. Furthermore, we elaborate on potential pitfalls and remaining challenges of these methods. Finally, we provide a brief outlook for the field as we foresee increasing adaptation and implementation of large scale data-driven approaches in material discovery and design campaigns.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Towards automating structural discovery in scanning transmission electron microscopy *

Abstract Scanning transmission electron microscopy is now the primary tool for exploring functional materials on the atomic level. Often, features of interest are highly localized in specific regions in the material, such as ferroelectric domain walls, extended defects, or second phase inclusions. Selecting regions to image for structural and chemical discovery via atomically resolved imaging has traditionally proceeded via human operators making semi-informed judgements on sampling locations and parameters. Recent efforts at automation for structural and physical discovery have pointed towards the use of ‘active learning’ methods that utilize Bayesian optimization with surrogate models to quickly find relevant regions of interest. Yet despite the potential importance of this direction, there is a general lack of certainty in selecting relevant control algorithms and how to balance a priori knowledge of the material system with knowledge derived during experimentation. Here we address this gap by developing the automated experiment workflows with several combinations to both illustrate the effects of these choices and demonstrate the tradeoffs associated with each in terms of accuracy, robustness, and susceptibility to hyperparameters for structural discovery. We discuss possible methods to build descriptors using the raw image data and deep learning based semantic segmentation, as well as the implementation of variational autoencoder based representation. Furthermore, each workflow is applied to a range of feature sizes including NiO pillars within a La:SrMnO 3 matrix, ferroelectric domains in BiFeO 3 , and topological defects in graphene. The code developed in this manuscript is open sourced and will be released at github.com/nccreang/AE_Workflows .

47 OTHER INSTRUMENTATION↗
DOE OSTI2025

Explainable and Differentiable Reinforcement Learning for Multi-objective Optimization in Particle Accelerators

Operating particle accelerators involves optimizing multiple goals simultaneously, which can be challenging due to trade-offs among objectives. While evolutionary algorithms like the genetic algorithm (GA) have been used for various Multi-Objective Optimization (MOO) tasks, they are not inherently suited for complex control problems. This talk highlights two variations of Reinforcement Learning (RL) for concurrently optimizing heat load and trip rates at the Continuous Electron Beam Accelerator Facility (CEBAF). The problem involves strict constraints on individual states, actions, and overall energy requirements of the beam. First, this talk highlights how differentiability can be harnessed through a Deep Differentiable Reinforcement Learning (DDRL) approach to address MOO issues within particle accelerators. We examine the DDRL method alongside Model Free Reinforcement Learning (MFRL), GA, and Bayesian Optimization (BO). The performance of these methods is assessed by generating a Pareto-front for two objectives. Our findings indicate that DDRL excels in handling high-dimensional problems more effectively than MFRL, BO, and GA. Next, we will show integration of explainable physics-based constraints into RL algorithms to enhance trans- parency and trust in decision-making processes by enabling users to verify that agents adhere to established physical principles. This surrogate function can be modeled using neural networks or sparse dictionary mod- els. By examining the mathematical form of the learned constraint function, we are able to confirm the agent has learned to use the established physics of each environment provided but the surrogate model. In addi- tion, we find that the introduction of a mathematical functional dictionary based surrogate model enables our reinforcement learning algorithms to reliably converge for difficult high-dimensional accelerator controls environments.

Rajput, Kishansingh [Thomas Jefferson National Acc↗
DOE OSTI2021

Efficient Active Learning for Gaussian Process Classification by Error Reduction

Active learning sequentially selects the best instance for labeling by optimizing an acquisition function to enhance data/label efficiency. The selection can be either from a discrete instance set (pool-based scenario) or a continuous instance space (query synthesis scenario). In this work, we study both active learning scenarios for Gaussian Process Classification (GPC). The existing active learning strategies that maximize the Estimated Error Reduction (EER) aim at reducing the classification error after training with the new acquired instance in a onestep-look-ahead manner. The computation of EER-based acquisition functions is typically prohibitive as it requires retraining the GPC with every new query. Moreover, as the EER is not smooth, it can not be combined with gradient-based optimization techniques to efficiently explore the continuous instance space for query synthesis. To overcome these critical limitations, we develop computationally efficient algorithms for EER-based active learning with GPC. Further, we derive the joint predictive distribution of label pairs as a one-dimensional integral, as a result of which the computation of the acquisition function avoids retraining the GPC for each query, remarkably reducing the computational overhead. We also derive the gradient chain rule to efficiently calculate the gradient of the acquisition function, which leads to the first query synthesis active learning algorithm implementing EER-based strategies. Our experiments clearly demonstrate the computational efficiency of the proposed algorithms. We also benchmark our algorithms on both synthetic and real-world datasets, which show superior performance in terms of sampling efficiency compared to the existing state-of-the-art algorithms.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

SANE: strategic autonomous non-smooth exploration for multiple optima discovery in multi-modal and non-differentiable black-box functions

Both computational and experimental material discovery bring forth the challenge of exploring multidimensional and multimodal parameter spaces, such as phase diagrams of Hamiltonians with multiple interactions, composition spaces of combinatorial libraries, material structure image spaces, and molecular embedding spaces. Often these systems are black-boxes and time-consuming to evaluate, which resulted in strong interest towards active learning methods such as Bayesian optimization (BO). However, these systems are often noisy which make the black box function severely multi-modal and non-differentiable, where a vanilla BO can get overly focused near a single or faux optimum, deviating from the broader goal of scientific discovery. To address these limitations, here we developed Strategic Autonomous Non-Smooth Exploration (SANE) to facilitate an intelligent Bayesian optimized navigation with a proposed cost-driven probabilistic acquisition function to find multiple global and local optimal regions, avoiding the tendency to becoming trapped in a single optimum. To distinguish between a true and false optimal region due to noisy experimental measurements, a human (domain) knowledge driven dynamic surrogate gate is integrated with SANE. We implemented the gate-SANE into pre-acquired piezoresponse spectroscopy data of a ferroelectric combinatorial library with high noise levels in specific regions, and piezoresponse force microscopy (PFM) hyperspectral data. SANE demonstrated better performance than classical BO to facilitate the exploration of multiple optimal regions and thereby prioritized learning with higher coverage of scientific values in autonomous experiments. Our work showcases the potential application of this method to real-world experiments, where such combined strategic and human intervening approaches can be critical to unlocking new discoveries in autonomous research.

Biswas, Arpan [University of Tennessee, Knoxville,↗
DOE OSTI2021

Bayesian inference of heterogeneous epidemic models: Application to COVID-19 spread accounting for long-term care facilities

Here we propose a high dimensional Bayesian inference framework for learning heterogeneous dynamics of a COVID-19 model, with a specific application to the dynamics and severity of COVID-19 inside and outside long-term care (LTC) facilities. We develop a heterogeneous compartmental model that accounts for the heterogeneity of the time-varying spread and severity of COVID-19 inside and outside LTC facilities, which is characterized by time-dependent stochastic processes and time-independent parameters in ~ 1500 dimensions after discretization. To infer these parameters, we use reported data on the number of confirmed, hospitalized, and deceased cases with suitable post-processing in both a deterministic inversion approach with appropriate regularization as a first step, followed by Bayesian inversion with proper prior distributions. To address the curse of dimensionality and the ill-posedness of the high-dimensional inference problem, we propose use of a dimension-independent projected Stein variational gradient descent method, and demonstrate the intrinsic low-dimensionality of the inverse problem. We present inference results with quantified uncertainties for both New Jersey and Texas, which experienced different epidemic phases and patterns. Moreover, we also present forecasting and validation results based on the empirical posterior samples of our inference for the future trajectory of COVID-19.

42 ENGINEERING↗
DOE OSTI2022

Functional Data Analysis for Extracting the Intrinsic Dimensionality of Spectra: Application to Chemical Homogeneity in the Open Cluster M67

High-resolution spectroscopic surveys of the Milky Way have entered the Big Data regime and have opened avenues for solving outstanding questions in Galactic archeology. However, exploiting their full potential is limited by complex systematics, whose characterization has not received much attention in modern spectroscopic analyses. In this work, we present a novel method to disentangle the component of spectral data space intrinsic to the stars from that due to systematics. Using functional principal component analysis on a sample of 18,933 giant spectra from APOGEE, we find that the intrinsic structure above the level of observational uncertainties requires ≈10 functional principal components (FPCs). Our FPCs can reduce the dimensionality of spectra, remove systematics, and impute masked wavelengths, thereby enabling accurate studies of stellar populations. To demonstrate the applicability of our FPCs, we use them to infer stellar parameters and abundances of 28 giants in the open cluster M67. We employ Sequential Neural Likelihood, a simulation-based Bayesian inference method that learns likelihood functions using neural density estimators, to incorporate non-Gaussian effects in spectral likelihoods. By hierarchically combining the inferred abundances, we limit the spread of the following elements in M67: Fe ≲ 0.02 dex; C ≲ 0.03 dex; O, Mg, Si, Ni ≲ 0.04 dex; Ca ≲ 0.05 dex; N, Al ≲ 0.07 dex (at 68% confidence). Our constraints suggest a lack of self-pollution by core-collapse supernovae in M67, which has promising implications for the future of chemical tagging to understand the star formation history and dynamical evolution of the Milky Way.

79 ASTRONOMY AND ASTROPHYSICS↗
NASA NTRS2011

Towards an Automated Classification of Transient Events in Synoptic Sky Surveys

We describe the development of a system for an automated, iterative, real-time classification of transient events discovered in synoptic sky surveys. The system under development incorporates a number of Machine Learning techniques, mostly using Bayesian approaches, due to the sparse nature, heterogeneity, and variable incompleteness of the available data. The classifications are improved iteratively as the new measurements are obtained. One novel featrue is the development of an automated follow-up recommendation engine, that suggest those measruements that would be the most advantageous in terms of resolving classification ambiguities and/or characterization of the astrophysically most interesting objects, given a set of available follow-up assets and their cost funcations. This illustrates the symbiotic relationship of astronomy and applied computer science through the emerging disciplne of AstroInformatics.

classification↗
DOE OSTI2023

Artificial intelligence for materials research at extremes

Abstract Materials development is slow and expensive, taking decades from inception to fielding. For materials research at extremes, the situation is even more demanding, as the desired property combinations such as strength and oxidation resistance can have complex interactions. Here, we explore the role of AI and autonomous experimentation (AE) in the process of understanding and developing materials for extreme and coupled environments. AI is important in understanding materials under extremes due to the highly demanding and unique cases these environments represent. Materials are pushed to their limits in ways that, for example, equilibrium phase diagrams cannot describe. Often, multiple physical phenomena compete to determine the material response. Further, validation is often difficult or impossible. AI can help bridge these gaps, providing heuristic but valuable links between materials properties and performance under extreme conditions. We explore the potential advantages of AE along with decision strategies. In particular, we consider the problem of deciding between low-fidelity, inexpensive experiments and high-fidelity, expensive experiments. The cost of experiments is described in terms of the speed and throughput of automated experiments, contrasted with the human resources needed to execute manual experiments. We also consider the cost and benefits of modeling and simulation to further materials understanding, along with characterization of materials under extreme environments in the AE loop. Graphical abstract AI sequential decision-making methods for materials research: Active learning, which focuses on exploration by sampling uncertain regions, Bayesian and bandit optimization as well as reinforcement learning (RL), which trades off exploration of uncertain regions with exploitation of optimum function value. Bayesian and bandit optimization focus on finding the optimal value of the function at each step or cumulatively over the entire steps, respectively, whereas RL considers cumulative value of the labeling function, where the latter can change depending on the state of the system (blue, orange, or green).

36 MATERIALS SCIENCE↗
DOE OSTI2024

AEOLUS: Advances in Experimental Design, Optimal Control, and Learning for Uncertain Complex Systems

Sustained advances in the mathematics of modeling and simulation have resulted in the capability today for routine simulation of a number of large scale complex DOE-relevant systems. As remarkable as this capability for solving the so-called forward problem is, it is typically only the first step-an inner loop within an outer loop that explores the simulation model's parameter space and decision space to characterize uncertainty in the model's predictions, learn unknown model parameters from data, design the most informative experiments, determine optimal control strategies, and create optimal designs. Broadly, what unifies all of these outer loop problems is that they are, in one form or another, optimization problems over parameter/control/design space that are constrained by complex uncertain models. To fully realize the power of scientific simulation as a basis for scientific discovery, technological innovation, and rational decision-making, it is imperative to move beyond simulation to tackle the outer loop of optimization for learning from data, experimental design, and control with complex uncertain models. When the models under consideration are large-scale and complex, and when the optimization variable and uncertain parameter spaces are high (or infinite) dimensional, this constitutes a grand challenge of the highest order, and is intractable with conventional methods. To overcome these challenges, the AEOLUS Center was established to develop a unified mathematical, computational, and statistical framework for (1) Learning predictive models from complex data via Bayesian inference and optimization, and (2) Optimizing experiments, processes, and designs using the resulting uncertain models. These problems are intractable with conventional methods, for several reasons: (1) The simulation problems that govern the inner loops of the optimization problems are expensive to execute (due to severe nonlinearity, heterogeneity, multiphysics/multiscale coupling); (2) The optimization variable and uncertain parameter spaces are high dimensional, often stemming from discretizations of infinite dimensional fields such as initial conditions, sources, or material properties. We argue that the key to overcoming these challenges is to develop new mathematical, computational, and statistical methods that exploit the structure of the Bayesian inference and optimization problems mediated by their underlying complex uncertain models. This structure includes the regularity, sparsity, geometry, low intrinsic dimensionality, and multifidelity nature of the maps from uncertain parameter/optimization variable spaces to the specific objectives targeted: Bayesian inference, optimal experimental design, and optimal control design. Black box methods developed as generic tools are incapable of exploiting this structure. To be successful, we must create, integrate, and cross-fertilize ideas across multiple areas of applied math--including approximation theory, Bayesian inference, data science, experimental design, information theory, machine learning, model reduction, optimal control theory, parallel algorithms, PDE-constrained optimization, randomized algorithms, stochastic optimization, and uncertainty quantification--all while exploiting the structure of the problems at hand. With this goal in mind, we have marshaled a team of leading authorities in these areas. While the methods we develop will be broadly applicable across a wide spectrum of DOE problems in which experiments inform models and the systems those models describe must be optimized under uncertainty, we have chosen a specific area, advanced manufacturing and materials, to drive our work. AMM is characterized by complex models across multiple scales, and is a rich source of challenging problems in inference, experimental design, and optimal control, requiring multifaceted and integrated advances in applied mathematics. As such, AMM serves as an excellent vehicle to motivate and demonstrate the advances in applied mathematics developed by our center.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS2022

A Framework for Deep Learning Emulation of Numerical Models With a Case Study in Satellite Remote Sensing

Numerical models based on physics represent the state of the art in Earth system modeling and comprise our best tools for generating insights and predictions. Despite rapid growth in computational power, the perceived need for higher model resolutions overwhelms the latest generation computers, reducing the ability of modelers to generate simulations for understanding parameter sensitivities and characterizing variability and uncertainty. Thus, surrogate models are often developed to capture the essential attributes of the full-blown numerical models. Recent successes of machine learning methods, especially deep learning (DL), across many disciplines offer the possibility that complex nonlinear connectionist representations may be able to capture the underlying complex structures and nonlinear processes in Earth systems. A difficult test for DL-based emulation, which refers to function approximation of numerical models, is to understand whether they can be comparable to traditional forms of surrogate models in terms of computational efficiency while simultaneously reproducing model results in a credible manner. A DL emulation that passes this test may be expected to perform even better than simple models with respect to capturing complex processes and spatiotemporal dependencies. Here, we examine, with a case study in satellite-based remote sensing, the hypothesis that DL approaches can credibly represent the simulations from a surrogate model with comparable computational efficiency. Our results are encouraging in that the DL emulation reproduces the results with acceptable accuracy and often even faster performance. We discuss the broader implications of our results in light of the pace of improvements in high-performance implementations of DL and the growing desire for higher resolution simulations in the Earth sciences.

Bayesian Deep Learning↗
DOE OSTI2015

Principles of Bayesian Ensemble Creation: Theoretical and Empirical Observations

Bayesian methods are theoretically optimal given their underlying assumptions. Bayesian model averaging (BMA) is generally considered the standard model for creating ensembles of learners using Bayesian methods, but this technique is often out-performed by more ad hoc methods in empirical studies. It has been proposed that Bayesian model averaging struggles in practice because it accounts for uncertainty about which model is the data-generating model but still operates under the assumption that only one of them is. Thus, BMA is unable to enrich the hypothesis space as effectively as other ensemble creation methods, serving as a form of optimal model selection, not optimal model combination. This work provides empirical verification for this hypothesis using both synthetic and real-world data. We propose several different approaches for Bayesian model combination, and demonstrate that they overcome this problem with BMA. We empirically show that even the most simplistic of Bayesian model combination strategies can compete with the traditional ad hoc techniques of bagging, boosting and stacking, as well as significantly outperforming BMA over a wide variety of cases. We also explore the implications of this observation as it relates to the effectiveness of ensemble learning.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Bayesian sequential optimal experimental design for nonlinear models using policy gradient reinforcement learning

We present a mathematical framework and computational methods for optimally designing a finite sequence of experiments. This sequential optimal experimental design (sOED) problem is formulated as a finite-horizon partially observable Markov decision process (POMDP) under a Bayesian setting and with information-theoretic utilities. The formulation is general and may accommodate continuous random variables, non-Gaussian posteriors, and nonlinear forward models. The sOED design policy incorporates elements of feedback and lookahead simultaneously, and we show it to generalize the commonly-used batch and greedy design strategies. We solve for the sOED policy using the policy gradient (PG) method from reinforcement learning, and provide a derivation for the PG expression in the sOED context. Adopting an actor-critic approach, the policy and value functions are parameterized using deep neural networks and improved via PG estimates produced from simulated episodes of designs and observations. The new PG-sOED algorithm is first validated on a linear-Gaussian benchmark, and then compared against other design baselines on a sensor movement problem for contaminant source inversion in a convection-diffusion field. As a result, we provide explanation for the policy behaviors using knowledge of the underlying physical process.

97 MATHEMATICS AND COMPUTING↗