Grid Search Approach Research Articles

Accuracy improvement in air‐quality forecasting using regressor combination with missing data imputation

AbstractThis article proposes a hybrid model based on regressor combination to improve the accuracy of air‐quality forecasting. The expectation‐maximization algorithm was used to impute the missing values of the dataset. The optimal hyperparameter values for the regressors were found by the grid search approach, depending on the mean absolute error (MAE), in the training session. The regressors having the minimum MAE were then globally combined for prediction. The output of the regressor with the minimum absolute error between the actual and predicted values was chosen as the prediction result of the hybrid model. The performance of the proposed model was compared with that of sequential deep learning methods, namely long short‐term memory and gated recurrent unit, in terms of MAE, mean relative error (MRE), and squared correlation coefficient (SCC) metrics. The imputed dataset was divided into training and testing subsets of different durations. According to the experimental results, our hybrid model performed better than the deep learning methods in terms of MAE, MRE, and SCC metrics, irrespective of the training data length. Furthermore, the Akaike's information criterion and the Bayesian information criterion values suggested that the quality of the hybrid model was better than that of the deep learning models.

Abstract 3027: Application of machine learning and grid search approaches to minimize lucitanib pharmacokinetic variability following different dosing regimens

Abstract Background: Lucitanib is a potent tyrosine kinase inhibitor that selectively inhibits vascular endothelial growth factor receptors 1, 2, and 3 (VEGFR1-3), platelet-derived growth factor receptors alpha and beta (PDGFRα/β), and fibroblast growth factor receptors types 1, 2, and 3 (FGFR1-3). Lucitanib is being evaluated in patients with solid tumors in combination with other anticancer agents. The pharmacokinetics (PK) of lucitanib were described by a 2-compartment population PK model with first-order absorption and elimination. Body weight was a covariate on clearance and volume of distribution. Objective: We sought to compare the variability of lucitanib PK between body weight-based and fixed-dose regimens to inform dosing regimen selection. Methods: Lucitanib PK were simulated for 5000 virtual patients with a uniform distribution of body weight (40-120 kg). Patients either received a fixed dose (10 mg once daily) or a body weight-based dose. Genetic and grid search algorithms were used to identify the body weight-based dosing regimens that resulted in the smallest fraction of outliers, defined as subjects outside of the 5th and 95th prediction interval of the area under the curve during the dosing interval (tau) at steady state (AUCtau,ss) for 70-kg patients receiving the fixed dose. The genetic algorithm used a continuous body weight and dose search space, whereas the grid search algorithm limited the search space to discrete values of body weight and dose (ie, 5 kg increments for body weight and 0.25 mg increments for dose). The penalty function in both algorithms was the fraction of outliers. The optimized body weight-based regimens were compared to the fixed-dose regimen to evaluate lucitanib PK variability. Results: The genetic and grid search algorithms identified similar and reasonable dosing regimens. Two or 3 body weight-based dosing regimens decreased the fraction of outliers by 1%-4% in the overall population (40-120 kg) versus the fixed dose. However, within the low body weight group (40-50 kg), body weight-based dosing regimens substantially decreased the fraction of outliers, with exposure greater than the 95th percentile reduced by 10%-15% compared with the fixed dose. On the other hand, the fixed-dose regimen predicted fewer outliers with exposure below the 5th percentile versus the body weight-based dosing regimens. Conclusions: The genetic algorithm provided similar results to the more computationally demanding method of grid searching. Compared with a fixed-dose regimen, body weight-based regimens showed a small reduction in PK variability for lucitanib in the overall population, suggesting that there is limited benefit with body weight-based dosing. Dose optimization will be further evaluated with a growing understanding of the target therapeutic range and exposure-response relationships. Citation Format: Michelle Liao, Jessie Zhou, Mark Sale, Jim J. Xiao. Application of machine learning and grid search approaches to minimize lucitanib pharmacokinetic variability following different dosing regimens [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 3027.

Optimizing deep learning model selection for angular feature extraction in satellite imagery

Deep learning techniques have been leveraged in numerous applications and across different data modalities over the past few decades, more recently in the domain of remotely sensed imagery. Given the complexity and depth of convolutional neural network (CNN) architectures, it is difficult to fully evaluate performance, optimize the hyperparameters, and provide robust solutions to a specific machine learning problem that can be applied to nontraditional real-world feature extraction and automation tasks. Ursa Space Systems Inc. develops machine learning approaches to build custom solutions and extract answers from synthetic aperture radar satellite data fused with other remote sensing data sets. One application is identifying the orientation of nontexture linear features in imagery, such as an inlet pipe on top of a cylindrical oil storage tank. We propose a two-phase approach for determining this orientation: first an optimized CNN is used in a nontraditional way to probabilistically determine a coarse location and orientation of the inlet pipe, followed by a maximum likelihood voting scheme to automatically extract the orientation of the angular feature within 7.5 deg. We use a known hyperparameter optimization technique to determine the best deep learning CNN architecture for our specific problem and under user-defined optimization and accuracy constraints, by optimizing model hyperparameters (number of layers, size of the input image, and data set preprocessing) using a manual and grid search approach. The use of this systematic approach for hyperparameter optimization yields increased accuracy for our angular feature extraction and orientation finding algorithm from 86% to 94%. Additionally, this proposed algorithm shows how machine learning can be used to improve real-world remote sensing workflows.

Determining the steering direction in critical situations: A decision tree–based method

Objective: Factors related to the driver–vehicle–environment system have a significant influence on a driver’s decision to perform evasive maneuvers, especially the decision of steering direction (DSD) in critical situations. However, few studies have systematically investigated the relationships between these factors and DSD. The objective of this study is to analyze and model drivers’ DSD in critical situations.Methods: Data from the NASS-CDS from 1995 to 2015 were utilized in this study. The decision tree (DT) classifier was utilized to model a driver’s DSD for both intersection-related and non-intersection-related subsets, combined with a 10-fold cross-validation technique and grid search approach to evaluate and optimize the model. An analysis of variable importance was also conducted.Results: Two separate DT models of drivers’ DSD were obtained based on the optimized hyperparameters, with test accuracies of 84.6% (intersection-related) and 79.2% (non-intersection-related). The variable DIFFANGLE (angle difference between 2 vehicles) ranked as the most important factor influencing drivers’ DSD in both models. The variables, in order of importance, were SPEED (travel speed of the subject vehicle) and AGE (driver’s age) for the intersection-related model and SPEED, PREMOVE (pre-event movement), TRAFFLOW (trafficway flow), and AGE for the non-intersection-related model. Moreover, an interesting same direction pattern was observed in both DT models.Conclusions: This study employed NASS-CDS data and DT classifiers to analyze and model drivers’ DSD behavior. The test accuracies for both classifiers were acceptable. Potential variables influencing drivers’ DSD were explored, which improves the research on evasive behavior in lateral movement and promotes further applications for intelligent vehicles using the constructed models.

7 Artificial Intelligence and Machine Learning as a Predictive Tool for Acute Kidney Injury in Trauma and Severely Burned Patients

Abstract Introduction Kidney injury doubles burn mortality—thus, early prediction of acute kidney injury (AKI) in the burn population could benefit from artificial intelligence (AI) and machine learning (ML). Our objective in this study was to build and assess the theoretical performances of such AI/ML algorithms and to develop generalizable models that could augment AKI recognition. Methods Two databases containing patients that received neutrophil gelatinase associated lipocalin (NGAL), creatinine, N-terminal pro-B-type natriuretic peptide (NT-proBNP) and urine output (UOP) measurements at admission were used to train, test, and generalize the AI/ML models. Models were first optimized in Cohort A for predicting AKI in Cohort B. Cohort A (n = 50) was based on a retrospective dataset of adult (age³18 years) burn patients, while Cohort B (n = 51) consisted of prospectively enrolled adult burned or non-burned trauma patients at risk for AKI. We employed a grid search and cross validation approach in building 68,100 unique ML models from five distinct ML approaches: logistic regression (LR), k-nearest neighbor (k-NN), support vector machine (SVM), random forest (RF), and deep neural networks (DNN) which enabled us to find the most accurate ML models. Results The best generalization accuracy (86%), sensitivity (91%), and specificity (85%) with NGAL alone was noted with LR, SVM and RF models. Generalizability prediction accuracy, sensitivity and specificity were respectively highest with the optimized DNN model (92%, 100%, and 90%) and the k-NN model (92%, 91%, and 93%) when tested with Cohort B using all four biomarkers. k-NN provided best generalization accuracy (84%) without NGAL using only NT-proBNP and creatinine, followed by DNN using creatinine only with an accuracy of 82%. AI/ML algorithms using results obtained at admission accelerated the average (SD) time to AKI prediction by 61.8 (32.5) hours. Conclusions NGAL is analytically superior to traditional AKI biomarkers such as creatinine and UOP. With machine learning, the AKI predictive capability of NGAL can be further enhanced and accelerated when combined with NT-proBNP, UOP, and creatinine. Applicability of Research to Practice Without NGAL, machine learning models continue to provide robust means in accelerating the prediction of AKI using both common and biomarkers of cardiorenal dysfunction.

A Method for Estimating the Crustal Azimuthal Anisotropy and Moho Orientation Simultaneously Using Receiver Functions

AbstractThe P‐to‐S converted phase at the Moho (the Pms phase) contains important information regarding the crustal structure. Here, we employ a grid search approach to simultaneously extract the bulk crustal azimuthal anisotropy and Moho orientation using receiver functions. The solution of this approach is the weighted average of two individual subfunctions that separately analyze the birefringence and arrival time of the Pms phase. Since a tilted Moho deflects the polarized components of the Pms phase, the Pms birefringence subfunction minimizes the energy on all the transverse polarization components (i.e., the transverse polarization direction before Pms splitting). The implementation of this subfunction is similar to that of the transverse component minimization method, but the Moho orientation is considered. The Pms arrival time subfunction takes into account both the crustal anisotropy and the Moho orientation to predict and fit Pms arrivals. According to extensive synthetic experiments, this approach is remarkably robust to the presence of strong anisotropy, a steep Moho, and slightly tilted fast axes (less than 30°). This approach is also highly tolerant of random noise and data with uneven azimuthal coverage. We apply this approach to three permanent stations on the Tibetan Plateau. The resulting measurements boast a high quality and are consistent with the crustal anisotropy and Moho orientation revealed in previous studies.

An Evaluation of Different Fast Fourier Transform - Transfer Learning Pipelines for the Classification of Wink-based EEG Signals

Brain Computer-Interfaces (BCI) offers a means of controlling prostheses for neurological disorder patients, primarily owing to their inability to control such devices due to their inherent physical limitations. More often than not, the control of such devices exploits the use of Electroencephalogram (EEG) signals. Nonetheless, it is worth noting that the extraction of the features is often a laborious undertaking. The use of Transfer Learning (TL) has been demonstrated to be able to mitigate the issue. However, the employment of such a method towards BCI applications, particularly with regards to EEG signals are limited. The present study aims to assess the effectiveness of a number of DenseNet TL models, viz. DenseNet169, DenseNet121 and DenseNet201 in extracting features for the classification of wink-based EEG signals. The extracted features are then classified through an optimised Random Forest (RF) classifier. The raw EEG signals are transformed into a spectrogram image via Fast Fourier Transform (FFT) before it was fed into selected TL models. The dataset was split with a stratified ratio of 60:20:20 into train, test, and validation datasets, respectively. The hyperparameters of the RF model was optimised through the grid search approach that utilises the five-fold cross-validation technique. It was established from the study that amongst the DenseNet pipelines evaluated, the DenseNet169 performed the best with an overall validation and test accuracy of 89%. The findings of the present investigation could facilitate BCI applications, e.g., for a grasping exoskeleton.

Improving Software Development effort estimating using Support Vector Regression and Feature Selection

Accurate and reliable software development effort estimation (SDEE) is one of the main concerns for project managers. Planning and scheduling a software projects using and inaccurate estimate may cause severe risks to software project under development such as delayed delivery, poor quality software, missing features. Therefore, accurate prediction of software effort plays an important role in the minimization of these risks that can lead to projects failure. Nowadays, application of artificial intelligence techniques has grown dramatically for predicting software effort. The researchers found that these techniques are suitable tools for accurate prediction. In this study, an improved model is designed for estimating software effort using support vector regression (SVR) and two feature selection (FS) methods. Prior to building model step, a preprocessing stage is performed by random forest or Boruta feature selection methods to remove unimportant features. Next, the SVR model is tuned by a grid search approach. The performance of model is then evaluated over eight well-known datasets through 30%holdout validation method. To show the impact of feature selection on the accuracy of SVR model, the proposed model was compared with SVR model without feature selection. The results indicated that SVR with feature selection outperforms SVR without FS in terms of the three accuracy measures used in this empirical study.

Model selection for metabolomics: predicting diagnosis of coronary artery disease using automated machine learning.

MotivationSelecting the optimal machine learning (ML) model for a given dataset is often challenging. Automated ML (AutoML) has emerged as a powerful tool for enabling the automatic selection of ML methods and parameter settings for the prediction of biomedical endpoints. Here, we apply the tree-based pipeline optimization tool (TPOT) to predict angiographic diagnoses of coronary artery disease (CAD). With TPOT, ML models are represented as expression trees and optimal pipelines discovered using a stochastic search method called genetic programing. We provide some guidelines for TPOT-based ML pipeline selection and optimization-based on various clinical phenotypes and high-throughput metabolic profiles in the Angiography and Genes Study (ANGES).ResultsWe analyzed nuclear magnetic resonance-derived lipoprotein and metabolite profiles in the ANGES cohort with a goal to identify the role of non-obstructive CAD patients in CAD diagnostics. We performed a comparative analysis of TPOT-generated ML pipelines with selected ML classifiers, optimized with a grid search approach, applied to two phenotypic CAD profiles. As a result, TPOT-generated ML pipelines that outperformed grid search optimized models across multiple performance metrics including balanced accuracy and area under the precision-recall curve. With the selected models, we demonstrated that the phenotypic profile that distinguishes non-obstructive CAD patients from no CAD patients is associated with higher precision, suggesting a discrepancy in the underlying processes between these phenotypes.Availability and implementationTPOT is freely available via http://epistasislab.github.io/tpot/.Supplementary information Supplementary data are available at Bioinformatics online.

Multi-Trait, Multi-Environment Genomic Prediction of Durum Wheat With Genomic Best Linear Unbiased Predictor and Deep Learning Methods.

Although durum wheat (Triticum turgidum var. durum Desf.) is a minor cereal crop representing just 5–7% of the world’s total wheat crop, it is a staple food in Mediterranean countries, where it is used to produce pasta, couscous, bulgur and bread. In this paper, we cover multi-trait prediction of grain yield (GY), days to heading (DH) and plant height (PH) of 270 durum wheat lines that were evaluated in 43 environments (country–location–year combinations) across a broad range of water regimes in the Mediterranean Basin and other locations. Multi-trait prediction analyses were performed by implementing a multi-trait deep learning model (MTDL) with a feed-forward network topology and a rectified linear unit activation function with a grid search approach for the selection of hyper-parameters. The results of the multi-trait deep learning method were also compared with univariate predictions of the genomic best linear unbiased predictor (GBLUP) method and the univariate counterpart of the multi-trait deep learning method (UDL). All models were implemented with and without the genotype × environment interaction term. We found that the best predictions were observed without the genotype × environment interaction term in the UDL and MTDL methods. However, under the GBLUP method, the best predictions were observed when the genotype × environment interaction term was taken into account. We also found that in general the best predictions were observed under the GBLUP model; however, the predictions of the MTDL were very similar to those of the GBLUP model. This result provides more evidence that the GBLUP model is a powerful approach for genomic prediction, but also that the deep learning method is a practical approach for predicting univariate and multivariate traits in the context of genomic selection.

New insights into seismic absorption imaging

In recent years, attenuation has been used as a marker for source and dynamic Earth processes due to its higher sensitivity to small variations of lithospheric properties compared to seismic velocity. From seismic hazard analysis to oil and gas exploration and rock physics, many fields need a better reconstruction of energy absorption, a constituent of seismic attenuation generally considered a reliable marker of fluid saturation in space. Here, we propose absorption tomography (AT), a technique grounded on the principles of scattering tomography and Multiple Lapse Time Window Analysis. We benchmark its efficiency to image absorption in space by comparing its results with those obtained using two of the most common mapping strategies, regionalization and kernel-based inversion for a diffusive regime. We applied these methodologies to three datasets, each characterized by a different tectonic setting and quality of the dataset: the Pollino fault area (Southern Italy), Mount St. Helens Volcano (USA) and Vrancea (Romania). AT overcomes the assignment of a single coda quality factor value between each source-receiver pair by modelling and inverting for the spatial distribution of energy as a function of different lapse times. It can then reconstruct node-dependent envelopes using analytic and computational sensitivity kernels and solve for coda attenuation with a grid-search approach. AT allows for a better reconstruction of the localized absorption anomalies than standard methodologies, even if the efficiency of the technique depends on the quality of the dataset, and identifies outliers in the data that could alter the final result and its interpretation. The implementation of analytical diffusive kernels in AT allows for a fast and highly-resolved imaging of well-sampled seismic faults structures. While slightly reducing resolution on faults and being computationally more expensive, implementing multiple-scattering lapse-time-dependent kernels still provides satisfactory results as well as a more physically-correct forward model.

Flood and Drought Prediction Using the Machine Learning Algorithm Support Vector Regression

Flood and drought are frequently happening natural disasters in most of the countries. These disasters can cause considerable damage to agriculture, ecology and economy of the country. Mitigating the impacts of flood and drought is a valuable help to the human being. The main cause of these disasters is precipitation. If the past precipitation data are analyzed properly, the future flood and drought events can be easily found. Prediction using the Standard Precipitation Index (SPI) is a way to find the wet or dry condition of a region or country. In this paper the SPI values with different lead times are calculated for a long period of time. These SPI indices are analysed by a predictive model using the machine learning algorithm called Support Vector Regression (SVR) with RBF (Radial Basis Function) kernel. In this model the Grid Search approach is used for optimization. The forecast result of this predictive model shows the predictive skill of the SVR-RBF kernel.

Measurement-Based Automatic Extraction of FET Parasitic Network by Linear Regression

A fully automatic procedure for the empirical extraction of field-effect transistors (FETs) parasitic network is proposed. By exploiting a grid search approach combined with linear regression, the identification requires Y-parameter measurements at a very few (even just one) bias points. The method is verified in simulation and applied to measurements of a 250-nm GaN high-electron-mobility transistor (HEMT).

Outpatient Readmission in Rheumatology: A Machine Learning Predictive Model of Patient's Return to the Clinic.

Our objective is to develop and validate a predictive model based on the random forest algorithm to estimate the readmission risk to an outpatient rheumatology clinic after discharge. We included patients from the Hospital Clínico San Carlos rheumatology outpatient clinic, from 1 April 2007 to 30 November 2016, and followed-up until 30 November 2017. Only readmissions between 2 and 12 months after the discharge were analyzed. Discharge episodes were chronologically split into training, validation, and test datasets. Clinical and demographic variables (diagnoses, treatments, quality of life (QoL), and comorbidities) were used as predictors. Models were developed in the training dataset, using a grid search approach, and performance was compared using the area under the receiver operating characteristic curve (AUC-ROC). A total of 18,662 discharge episodes were analyzed, out of which 2528 (13.5%) were followed by outpatient readmissions. Overall, 38,059 models were developed. AUC-ROC, sensitivity, and specificity of the reduced final model were 0.653, 0.385, and 0.794, respectively. The most important variables were related to follow-up duration, being prescribed with disease-modifying anti-rheumatic drugs and corticosteroids, being diagnosed with chronic polyarthritis, occupation, and QoL. We have developed a predictive model for outpatient readmission in a rheumatology setting. Identification of patients with higher risk can optimize the allocation of healthcare resources.

Three‐Dimensional Acoustic Multipole Waveform Inversion at Yasur Volcano, Vanuatu

AbstractAcoustic waveform inversions can provide estimates of volume flow rate and erupted mass, enhancing the ability to estimate volcanic emissions. Previous studies have generally assumed a simple acoustic source (monopole); however, more complex and accurate source reconstructions are possible with a combination of equivalent sources (multipole). We deployed a high‐density acoustic network around Yasur volcano, Vanuatu, including acoustic sensors on a tethered aerostat that was moved every ∼15–60 min. Using this unique data set we invert for the acoustic multipole source mechanism using a grid search approach for 80 events to examine volume flow rates and dipole strengths. Our method utilizes finite‐difference time‐domain modeling to obtain the full 3‐D Green's functions that account for topography. Inversion results are compared using a monopole‐only, multipole (monopole and dipole), simulations that do not include topography, and those that use a subset of sensors. We find that the monopole source is a good approximation when topography is considered. However, initial compression amplitude is not fully captured by a monopole source so source directionality cannot be ruled out. The monopole solution is stable regardless of whether a monopole‐only or multipole inversion is performed. Inversions for the dipole components produce estimates consistent with observed source directionality, though these inversions are somewhat unstable given station configurations of typical deployments. Our results suggest that infrasound waveform inversion shows promise for realistic quantitative source estimates, but additional work is necessary to fully explore inversion stability, uncertainty, and robustness.

A grid-search approach for 4D pressure-saturation discrimination

We develop a grid-search-based inversion method to discriminate between changes of pore pressure and water saturation. In this approach, we build a forward model by combining different rock-physics models and we predict grids of time-lapse attributes within a range of pressure and saturation. The time-lapse measurements from the data are matched with these predicted attributes to find the solution. We suggest using the independent time shifts and time-lapse attenuation to stabilize the inversion result. A synthetic example illustrates the advantages of the inversion workflow. The method helps to visualize how pressure and saturation change the elastic properties. This advantage is also useful in quality control and uncertainty analysis. We apply our method to a real underground blowout example in the North Sea area where gas has leaked into two sand layers at different conditions. We find that gas migrated according to the sand structure with the dipping layer leading to faster migration rates. The inversion results indicate that careful data preparation is crucial, not only to obtain stable inversion performance, but also to capture valuable information when interpreting the blowout development.

Vehicular traffic noise prediction and propagation modelling using neural networks and geospatial information system.

This study proposes a neural network (NN) model to predict and simulate the propagation of vehicular traffic noise in a dense residential area at the New Klang Valley Expressway (NKVE) in Shah Alam, Malaysia. The proposed model comprises of two main simulation steps: that is, the prediction of vehicular traffic noise using NN and the simulation of the propagation of traffic noise emission using a mathematical model. First, the NN model was developed with the following selected noise predictors: the number of motorbikes, the sum of vehicles, car ratio, heavy vehicle ratio (e.g. truck, lorry and bus), highway density and a light detection and ranging (LiDAR)-derived digital surface model (DSM). Subsequently, NN and its hyperparameters were optimised by a systematic optimisation procedure based on a grid search approach. The noise propagation model was then developed in a geographic information system (GIS) using five variables, namely road geometry, barriers, distance, interaction of air particles and weather parameters. The noise measurement was conducted continuously at 15-min intervals and the data were analysed by taking the minimum, maximum and average values recorded during the day. The measurement was performed four times a day (i.e. morning, afternoon, evening, and midnight) over two days of the week (i.e. Sunday and Monday). An optimal radial basis function NN was used with 17 hidden layers. The learning rate and momentum values were 0.05 and 0.9, respectively. Finally, the accuracy of the proposed method achieved 78.4% with less than 4.02dB (A) error in noise prediction. Overall, the proposed models were found to be promising tools for traffic noise assessment in dense urban areas.

Modeling of CO Emissions from Traffic Vehicles Using Artificial Neural Networks

Traffic emissions are considered one of the leading causes of environmental impact in megacities and their dangerous effects on human health. This paper presents a hybrid model based on data mining and GIS models designed to predict vehicular Carbon Monoxide (CO) emitted from traffic on the New Klang Valley Expressway, Malaysia. The hybrid model was developed based on the integration of GIS and the optimized Artificial Neural Network algorithm that combined with the Correlation based Feature Selection (CFS) algorithm to predict the daily vehicular CO emissions and generate prediction maps at a microscale level in a small urban area by using a field survey and open source data, which are the main contributions to this paper. The other contribution is related to the case study, which represents the spatial and quantitative variations in the vehicular CO emissions between toll plaza areas and road networks. The proposed hybrid model consists of three steps: the first step is the implementation of the correlation-based Feature Selection model to select the best model’s predictors; the second step is the prediction of vehicular CO by using a multilayer perceptron neural network model; and the third step is the creation of micro scale prediction maps. The model was developed using six traffic CO predictors: number of vehicles, number of heavy vehicles, number of motorbikes, temperature, wind speed and a digital surface model. The network architecture and its hyperparameters were optimized through a grid search approach. The traffic CO concentrations were observed at 15-min intervals on weekends and weekdays, four times per day. The results showed that the developed model had achieved validation accuracy of 80.6 %. Overall, the developed models are found to be promising tools for vehicular CO simulations in highly congested areas.

Efficient Selection of Hyperparameters in Large Bayesian VARs Using Automatic Differentiation

Large Bayesian VARs with the natural conjugate prior are now routinely used for forecasting and structural analysis. It has been shown that selecting the prior hyperparameters in a data-driven manner can often substantially improve forecast performance. We propose a computationally efficient method to obtain the optimal hyperparameters based on Automatic Differentiation, which is an efficient way to compute derivatives. Using a large US dataset, we show that using the optimal hyperparameter values leads to substantially better forecast performance. Moreover, the proposed method is much faster than the conventional grid-search approach, and is applicable in high-dimensional optimization problems. The new method thus provides a practical and systematic way to develop better shrinkage priors for forecasting in a data-rich environment.

Crash injury severity analysis using a two-layer Stacking framework

Crash injury severity analysis is useful for traffic management agency to further understand severity of crashes. A two-layer Stacking framework is proposed in this study to predict the crash injury severity: The fist layer integrates advantages of three base classification methods: RF (Random Forests), AdaBoost (Adaptive Boosting), and GBDT (Gradient Boosting Decision Tree); the second layer completes classification of crash injury severity based on a Logistic Regression model. A total of 5538 crashes were recorded at 326 freeway diverge areas. In the model calibration, several parameters including the number of trees in three base classification methods, learning rate, and regularization coefficient are optimized via a systematic grid search approach. In the model validation, the performance of the Stacking model is compared with several traditional models including the Support Vector Machine (SVM), Multi-Layer Perceptron (MLP) and Random Forests (RF) in the multi classification experiments. The prediction results show that Stacking model achieves superior performance evaluated by two indicators: accuracy and recall. Furthermore, all the factors used in severity prediction are classified into different categories according to their influence on the results, and sensitivity analysis of several significant factors is finally implemented to explore the impact of their value variation on the prediction accuracy.