Selecting linear models under the Bayesian paradigm with focus on good prediction over a user-specified distribution on the covariate space
Abstract Not Provided
SEARCH · Search NASA
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.
Abstract Not Provided
This paper compares three approaches for model selection: classical least squares methods, information theoretic criteria, and Bayesian approaches. Least squares methods are not model selection methods although one can select the model that yields the smallest sum-of-squared error function. Information theoretic approaches balance overfitting with model accuracy by incorporating terms that penalize more parameters with a log-likelihood term to reflect goodness of fit. Bayesian model selection involves calculating the posterior probability that each model is correct, given experimental data and prior probabilities that each model is correct. As part of this calculation, one often calibrates the parameters of each model and this is included in the Bayesian calculations. Our approach is demonstrated on a structural dynamics example with models for energy dissipation and peak force across a bolted joint. The three approaches are compared and the influence of the log-likelihood term in all approaches is discussed.
In this research, Bayesian model averaging (BMA) and least absolute shrinkage and selection operator regression (LASSO) algorithms were adopted for the core permeability modeling as a function of well log and core measurements. More specifically, the core permeability (dependent factor) was modeled given the well log and core data (independent variables) and then predicted in non-cored intervals of a well in a sandstone formation. The BMA is a stochastic linear modeling and a Bayesian parameter selection. Among 50 linear models generated as a function of the independent variables and depicted in Occam’s window, the best model of the highest posterior probability is determined. The best model has the optimal subset parameters that most influences the response factor. The core permeability modeling was again conducted using the LASSO algorithm, which adopts a completely different way of modeling and subset selection by using the penalized least-squared equation. Both BMA and LASSO resulted in a very accurate prediction of the core permeability by achieving perfect measured and calculated permeability matching. Matches for the two algorithms were quantified by computing the adjusted R-squared and root-mean-squared prediction error. In addition, results of the BMA and LASSO were compared to the conventional multiple linear regression (LM). The LASSO algorithm led to an accurate matching similar to LM, but slightly better than BMA. Therefore, both BMA and LASSO represent integrated procedures for accurate permeability estimation similar to conventional regression analysis.
A framework is presented to quantify, based on Bayesian evidence, the relative plausibility of species se- lection options in rate-controlled constrained equilibrium (RCCE) reduced chemical models, accounting for uncertainty in the kinetic parameters and experimental data used to refine them. This approach bal- ances the joint goals of matching available data and avoiding overfitting, which is well-understood to limit extrapolative capacity for true prediction. The methodology is applied to homogeneous autoigni- tion, where predictions are known to be particularly sensitive to chemical model details, specially at low temperatures. It is first introduced for hydrogen–air autoignition using an established mechanism, then demonstrated in two applications of methane–air autoignition using the larger GRI-1.2 mechanism. This larger mechanism significantly increases the computational cost of model selection (though not of the subsequent application in predictions), which is alleviated with a time-scale-guided pre-sorting strategy. Uses and extensions of this new formulation are discussed.
Computational investigations of heterogeneously catalyzed reactions using density functional theory (DFT) are often inaccurate, largely due to uncertainties in the choice of DFT functional (enthalpic uncertainty) and approximations for modeling adsorbate movement along the catalyst surface (entropic uncertainty). This work illustrates that both uncertainties are significant in the investigation of ethane dehydrogenation (EDH) and hydrogenolysis on Pt catalysts by considering the complete deconstruction of ethane on Pt(111), Pt(211), and Pt(100) using microkinetic modeling (MKM). Hence, this work uses both noncalibrated and Bayesian-calibrated MKMs to quantify and correct inaccuracies in macroscopic properties due to both uncertainties. A Bayesian approach to the correction of entropic errors was introduced using a “Modified Fermi Function (MFF)” to calibrate between the two bounds of entropy represented by the harmonic oscillator (HO) and free translator (FT) approximations. Regardless of enthalpic and entropic uncertainties, all three surfaces are capable of ethane activation; however, Pt(211) was found to be the most active and is largely responsible for methane production. Next, Pt(111) is largely responsible for acetylene production, and Pt(100) has the highest ethylene selectivity but is most susceptible to coking. By comparison of different calibrated models, the FT entropy approximation was found to better describe EDH under typical experimental conditions. Statistical evidence was found to support Pt(111) as the active site for EDH, assuming that one single site is responsible for the chemistry. On the three surfaces, competing second dehydrogenations to CH 2 CH 2 and CH 3 CH were observed as well as isomerization of CH 3 CH back to CH 2 CH 2 and deeper dehydrogenation of CH 3 CH. In conclusion, C–C cleavage was found to largely proceed via the CH 3 C intermediate on Pt(100) and Pt(111), while on Pt(211), it was via both CHC and CH 3 C.
Uncertainty analysis, reported experimental literature data, and density functional theory were synthesized to model the effect of surface tin coverage on platinum-based catalysts for nonoxidative propane dehydrogenation to propylene. Here, this study tests four different platinum–tin skin surface models as potential catalytic sites, Pt 3 Sn/Pt(100), PtSn/Pt(100), Pt 3 Sn/Pt(111), and Pt 2 Sn/Pt(211), and compares them to the corresponding pure Pt surface sites using an uncertainty analysis methodology that uses BEEF-vdW with its ensembles (BMwE) to generate the uncertainty for the energies of the intermediates and transition states. One experimental data set with two experimental observations, selectivity to propylene and turnover frequency of propylene, was used as a calibration data set to evaluate the impact of the experimental data on informing the models. This study finds that the prior model for Pt 3 Sn/Pt(100) is the most active and Pt 2 Sn/Pt(211) is the most selective toward propylene. Active sites on the (100) facet have the highest probability of being responsible for C 1 and C 2 product formations (C–C bond cleavage). Increasing the Sn coverage on the (100) surface facet to a PtSn/Pt(100) active site leads to a significantly reduced rate and might explain the experimentally observed higher selectivity of Sn-doped catalysts relative to pure Pt catalysts. Next, this study finds that for all surfaces, except PtSn/Pt(100), the rate-controlling steps are the initial dehydrogenation steps alongside some partially rate-controlling second dehydrogenation steps. For PtSn/Pt(100), only the initial terminal dehydrogenation step to CH 3 CH 2 CH 2 * and second dehydrogenation steps are rate-controlling. Next, the calibrated models for all surfaces were found to be selective toward propylene production and model the reported turnover frequency successfully. Nevertheless, Pt 2 Sn/Pt(211) emerges as the active site with some (minor) evidence as the main active site based on Jeffreys’ scale interpretation of Bayes factors. This observation agrees with prior studies that also found step sites to be most likely the most relevant active sites for pure Pt catalysts. Overall, the results indicate that tin, in addition to affecting the binding strength of the adsorbed species, prevents deeper dehydrogenation (reducing coking) and cracking reactions through increasing activation barriers for unwanted side reactions.
This paper addresses the issue of overfitting while calibrating unknown parameters of over-parameterized physics-based models with noisy and incomplete observations. Here, a semi-analytical Bayesian framework of nonlinear sparse Bayesian learning (NSBL) is proposed to identify sparsity among model parameters during Bayesian inversion. NSBL offers significant advantages over machine learning algorithm of sparse Bayesian learning (SBL) for physics-based models, such as 1) the likelihood function or the posterior parameter distribution is not required to be Gaussian, and 2) prior parameter knowledge is incorporated into sparse learning (i.e. not all parameters are treated as questionable). NSBL employs the concept of automatic relevance determination (ARD) to facilitate sparsity among questionable parameters through parameterized prior distributions. The analytical tractability of NSBL is enabled by employing Gaussian ARD priors and by building a Gaussian mixture-model approximation of the posterior parameter distribution that excludes the contribution of ARD priors. Subsequently, type-II maximum likelihood is executed using Newton's method whereby the evidence and its gradient and Hessian information are computed in a semi-analytical fashion. We show numerically and analytically that SBL is a special case of NSBL for linear regression models. Subsequently, a linear regression example involving multimodality in both parameter posterior pdf and model evidence is considered to demonstrate the performance of NSBL in cases where SBL is inapplicable. Next, NSBL is applied to identify sparsity among the damping coefficients of a mass-spring-damper model of a shear building frame. These numerical studies demonstrate the robustness and efficiency of NSBL in alleviating overfitting during Bayesian inversion of nonlinear physics-based models.
Using the temperature data from Planck we search for departures from a power-law primordial power spectrum, employing Bayesian model-selection and posterior probabilities. We parametrize the spectrum with n knots located at arbitrary values of logk, with both linear and cubic splines. This formulation recovers both slow modulations and sharp transitions in the primordial spectrum. The power spectrum is well-fit by a featureless, power-law at wavenumbers k>10{sup -3} Mpc{sup -1}. A modulated primordial spectrum yields a better fit relative to ΛCDM at large scales, but there is no strong evidence for a departure from a power-law spectrum. Moreover, using simulated maps we show that a local feature at k ∼ 10{sup -3} Mpc{sup -1} can mimic the suppression of large-scale power. With multi-knot spectra we see only small changes in the posterior distributions for the other free parameters in the standard ΛCDM universe. Lastly, we investigate whether the hemispherical power asymmetry is explained by independent features in the primordial power spectrum in each ecliptic hemisphere, but find no significant differences between them.
It has recently been demonstrated that, in the event of a putative signal in dark matter direct detection experiments, properly identifying the underlying dark matter-nuclei interaction promises to be a challenging task. Given the most optimistic expectations for the number counts of recoil events in the forthcoming Generation 2 experiments, differentiating between interactions that produce distinct features in the recoil energy spectra will only be possible if a strong signal is observed simultaneously on a variety of complementary targets. However, there is a wide range of viable theories that give rise to virtually identical energy spectra, and may only differ by the dependence of the recoil rate on the dark matter velocity. In this work, we investigate how degeneracy between such competing models may be broken by analyzing the time dependence of nuclear recoils, i.e. the annual modulation of the rate. For this purpose, we simulate dark matter events for a variety of interactions and experiments, and perform a Bayesian model-selection analysis on all simulated data sets, evaluating the chance of correctly identifying the input model for a given experimental setup. Lastly, we find that including information on the annual modulation of the rate may significantly enhance the ability of a single target to distinguish dark matter models with nearly degenerate recoil spectra, but only with exposures beyond the expectations of Generation 2 experiments.
We present a sample of galaxies with the Dark Energy Survey (DES) photometry that replicates the properties of the BOSS CMASS sample. The CMASS galaxy sample has been well characterized by the Sloan Digital Sky Survey (SDSS) collaboration and was used to obtain the most powerful redshift-space galaxy clustering measurements to date. A joint analysis of redshift-space distortions (such as those probed by CMASS from SDSS) and a galaxy-galaxy lensing measurement for an equivalent sample from DES can provide powerful cosmological constraints. Unfortunately, the DES and SDSS-BOSS footprints have only minimal overlap, primarily on the celestial equator near the SDSS Stripe 82 region. Using this overlap, we build a robust Bayesian model to select CMASS-like galaxies in the remainder of the DES footprint. The newly defined DES-CMASS (DMASS) sample consists of 117,293 effective galaxies covering $1,244 {\rm deg}^2$. Through various validation tests, we show that the DMASS sample selected by this model matches well with the BOSS CMASS sample, specifically in the South Galactic cap (SGC) region that includes Stripe 82. Combining measurements of the angular correlation function and the clustering-z distribution of DMASS, we constrain the difference in mean galaxy bias and mean redshift between the BOSS CMASS and DMASS samples to be $\Delta b = 0.010^{+0.045}_{-0.052}$ and $\Delta z = \left( 3.46^{+5.48}_{-5.55} \right) \times 10^{-3}$ for the SGC portion of CMASS, and $\Delta b = 0.044^{+0.044}_{-0.043} $ and $\Delta z= ( 3.51^{+4.93}_{-5.91}) \times 10^{-3}$ for the full CMASS sample. These values indicate that the mean bias of galaxies and mean redshift in the DMASS sample is consistent with both CMASS samples within $1\sigma$.
Not Available
Explore the source record for details and available documents.
A novel Bayesian approach significantly accelerates data collection for metal oxide reduction/re-oxidation thermodynamic fitting.
Model specification remains challenging in spectroscopy of plant biochemistry, as exemplified by the availability of various spectral indices or band combinations for estimating the same biochemical. This lack of consensus in model choice across applications argues for a paradigm shift in hyperspectral methods to address model uncertainty and misspecification. We demonstrated one such method using Bayesian model averaging (BMA), which performs variable/band selection and quantifies the relative merits of many candidate models to synthesize a weighted average model with improved predictive performances. The utility of BMA was examined using a portfolio of 27 foliage spectral–chemical datasets representing over 80 species across the globe to estimate multiple biochemical properties, including nitrogen, hydrogen, carbon, cellulose, lignin, chlorophyll (a or b), carotenoid, polar and nonpolar extractives, leaf mass per area, and equivalent water thickness. We also compared BMA with partial least squares (PLS) and stepwise multiple regression (SMR). Results showed that all the biochemicals except carotenoid were accurately estimated from hyerspectral data with R2 values > 0.80.
Abstract not provided.
Codon usage bias (CUB), the non-uniform usage of synonymous codons, occurs across all domains of life. Adaptive CUB is hypothesized to result from various selective pressures, including selection for efficient ribosome elongation, accurate translation, mRNA secondary structure, and/or protein folding. Given the critical link between protein folding and protein function, numerous studies have analyzed the relationship between codon usage and protein structure. The results from these studies have often been contradictory, likely reflecting the differing methods used for measuring codon usage and the failure to appropriately control for confounding factors, such as differences in amino acid usage between protein structures and changes in the frequency of different structures with gene expression. Here we take an explicit population genetics approach to quantify codon-specific shifts in natural selection related to protein structure in S. cerevisiae and E. coli. Unlike other metrics of codon usage, our approach explicitly separates the effects of natural selection, scaled by gene expression, and mutation bias while naturally accounting for a region’s amino acid usage. Bayesian model comparisons suggest selection on codon usage varies only slightly between helix, sheet, and coil secondary structures and, similarly, between structured and intrinsically-disordered regions. Similarly, in contrast to previous findings, we find selection on codon usage only varies slightly at the termini of helices in E. coli. Using simulated data, we show this previous work indicating “non-optimal” codons are enriched at the beginning of helices in S. cerevisiae was due to failure to control for various confounding factors (e.g. amino acid biases, gene expression, etc.), and rather than selection to modulate cotranslational folding. Our results reveal a weak relationship between codon usage and protein structure, indicating that differences in selection on codon usage between structures are slight. In addition to the magnitude of differences in selection between protein structures being slight, the observed shifts appear to be idiosyncratic and largely codon-specific rather than systematic reversals in the nature of selection. Overall, our work demonstrates the statistical power and benefits of studying selective shifts on codon usage or other genomic features from an explicitly evolutionary approach. Limitations of this approach and future potential research avenues are discussed.
Using ~100 X-ray selected clusters in the Dark Energy Survey Science Verification data, in this work we constrain the luminosity function (LF) of cluster red-sequence galaxies as a function of redshift. This is the first homogeneous optical/X-ray sample large enough to constrain the evolution of the LF simultaneously in redshift (0.1 < z < 1.05) and cluster mass ($13.5 \le \rm {log_{10}}(M_{200crit}) \sim \lt 15.0$). We pay particular attention to completeness issues and the detection limit of the galaxy sample. We then apply a hierarchical Bayesian model to fit the cluster galaxy LFs via a Schechter function, including its characteristic break (m*) to a faint end power-law slope (α). Our approach allows us to avoid known issues in similar analyses based on stacking or binning the clusters. We find weak and statistically insignificant (~1.9σ) evolution in the faint end slope α versus redshift. We also find no dependence in α or m* with the X-ray inferred cluster masses. Yet, the amplitude of the LF as a function of cluster mass is constrained to ${\sim } 20{{\ \rm per\ cent}}$ precision. As a by-product of our algorithm, we utilize the correlation between the LF and cluster mass to provide an improved estimate of the individual cluster masses as well as the scatter in true mass given the X-ray inferred masses. This technique can be applied to a larger sample of X-ray or optically selected clusters from the Dark Energy Survey, greatly improving the sensitivity of the analysis.