New Approach to Designing Multilayer Feedforward Neural Network Architecture for Modeling Nonlinear Restoring Forces. I: Formulation

This paper addresses the modeling problem of nonlinear and hysteretic dynamic behaviors through a constructive modeling approach which exploits existing mathematical concepts in artificial neural network modeling. In contrast with many neural network applications, which often result in large and complex “black-box” models, here, the writers strive to produce phenomenologically accurate model behavior starting with network architecture of manageable/small sizes. This affords the potential of creating relationships between model parameter values and observed phenomenological behaviors. Here a linear sum of basis functions is used in modeling nonlinear hysteretic restoring forces. In particular, nonlinear sigmoidal activation functions are chosen as the core building block for their robustness in approximating arbitrary functions. The appropriateness and effectiveness of this set of basis function in modeling a wide variety of nonlinear dynamic behaviors observed in structural mechanics are depicted from an algebraic and geometric perspective.     

Hsp70 and Hsc70 are preferentially expressed in differentiated epithelial cells in normal human endometrium and ectocervix

This paper addresses parametric system identification of linear and nonlinear dynamic systems by analysis of the input and output signals. Specifically, we investigate the relationship between estimation of the system using a feedforward neural network model and estimation of the system by use of linear and nonlinear autoregressive moving-average (ARMA) models. By utilizing a neural network model incorporating a polynomial activation function, we show the equivalence of the artificial neural network to the linear and nonlinear ARMA models. We compare the parameterization of the estimated system using the neural network and ARMA approaches by utilizing data generated by means of computer simulations. Specifically, we show that the parameters of a simulated ARMA system can be obtained from the neural network analysis of the simulated data or by conventional least squares ARMA analysis. The feasibility of applying neural networks with polynomial activation functions to the analysis of experimental data is explored by application to measurements of heart rate (HR) and instantaneous lung volume (ILV) fluctuations.     

Modeling methodology for nonlinear physiological systems

A general modeling approach for a broad class of nonlinear systems is presented that uses the concept of principal dynamic modes (PDMs). These PDMs constitute a filter bank whose outputs feed into a multi-input static nonlinearity of multinomial (polynomial) form to yield a general model for the broad class of Volterra systems. Because the practically obtainable models (from stimulus-response data) are of arbitrary order of nonlinearity, this approach is applicable to many nonlinear physiological systems heretofore beyond our methodological means. Two specific methods are proposed for the estimation of these PDMs and the associated nonlinearities from stimulus-response data. Method I uses eigendecomposition of a properly constructed matrix using the first two kernel estimates (obtained by existing methods). Method II uses a particular class of feedforward artificial neural networks with polynomial activation functions. The efficacy of these two methods is demonstrated with computer-simulated examples, and their relative performance is discussed. The advent of this approach promises a practicable solution to the vexing problem of modeling highly nonlinear physiological systems, provided that experimental data be available for reliable estimation of the requisite PDMs.     

Prediction of partition coefficient based on atom-type electrotopological state indices

The aim of this study was to determine the efficacy of atom-type electrotopological state indices for estimation of the octanol-water partition coefficient (log P) values in a set of 345 drug compounds or related complex chemical structures. Multilinear regression analysis and artificial neural networks were used to construct models based on molecular weights and atom-type electrotopological state indices. Both multilinear regression and artificial neural networks provide reliable log P estimations. For the same set of parameters, application of neural networks provided better prediction ability for training and test sets. The present study indicates that atom-type electrotopological state indices offer valuable parameters for fast evaluation of octanol-water partition coefficients that can be applied to screen large databases of chemical compounds, such as combinatorial libraries.     

Estimation of Aeroelastic Parameters of Bridge Decks Using Neural Networks

A new method of estimating flutter derivatives using artificial neural networks is proposed. Unlike other computational fluid dynamics based numerical analyses, the proposed method estimates flutter derivatives utilizing previously measured experimental data. One of the advantages of the neural networks approach is that they can approximate a function of many dimensions. An efficient method has been developed to quantify the geometry of deck sections for neural network input. The output of the neural network is flutter derivatives. The flutter derivatives estimation network, which has been trained by the proposed methodology, is tested both for training sets and novel testing sets. The network shows reasonable performance for the novel sets, as well as outstanding performance for the training sets. Two variations of the proposed network are also presented, along with their estimation capability. The paper shows the potential of applying neural networks to wind force approximations.     

Applying artificial neural network models to clinical decision making.

Because psychological assessment typically lacks biological gold standards, it traditionally has relied on clinicians' expert knowledge. A more empirically based approach frequently has applied linear models to data to derive meaningful constructs and appropriate measures. Statistical inferences are then used to assess the generality of the findings. This article introduces artificial neural networks (ANNs), flexible nonlinear modeling techniques that test a model's generality by applying its estimates against "future" data. ANNs have potential for overcoming some shortcomings of linear models. The basics of ANNs and their applications to psychological assessment are reviewed. Two examples of clinical decision making are described in which an ANN is compared with linear models, and the complexity of the network performance is examined. Issues salient to psychological assessment are addressed. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Risk Assessment for Bridge Maintenance Projects: Neural Networks versus Regression Techniques

Bridge risk assessment often serves as the basis for bridge maintenance priority ranking and optimization and conducted periodically for the purpose of safety. This paper presents an application of artificial neural networks in bridge risk assessment, in which back-propagation neural networks are developed to model bridge risk score and risk categories. The study investigated and utilized 506 bridge maintenance projects to develop the models. It is shown that neural networks have a very strong capability of modeling and classifying bridge risks. The average accuracies for risk score and risk categories are both over 96%. A comparative study is conducted with an alternative methodology using multiple regression techniques. The results revealed that neural networks achieved much better performances than regression analysis models. In addition an integrated forecasting approach was utilized to combine neural networks and regression analysis to generate hybrid models, which produced better accuracies than any of the individually developed models.     

Hybrid Models of Neural Networks and Genetic Algorithms for Predicting Preliminary Cost Estimates

This technical note applies hybrid models of neural networks (NN) and genetic algorithms (GA) to cost estimation of residential buildings to predict preliminary cost estimates. Data used in the study are for residential buildings constructed from 1997 to 2000 in Seoul, Korea. These are used in training each model and evaluating its performance. The models applied were Model I, which determines each parameter of a back-propagation network by a trial-and-error process; Model II, which determines each parameter of a back-propagation network by GAs; and Model III, which trains weights of NNs using genetic algorithms. The research revealed that optimizing each parameter of back-propagation networks using GAs is most effective in estimating the preliminary costs of residential buildings. Therefore, GAs may help estimators overcome the problem of the lack of adequate rules for determining the parameters of NNs.     

Stage–Discharge Relations for Low-Gradient Tidal Streams Using Data-Driven Models

Development of stage–discharge relationships for coastal low-gradient streams is a challenging task. Such relationships are highly nonlinear, nonunique, and often exhibit multiple loops. Conventional parametric regression methods usually fail to model these relationships. Therefore, this study examines the utility of two data-driven computationally intensive modeling techniques namely, artificial neural networks and local nonparametric regression, to model such complex relationships. The results show an overall good performance of both modeling techniques. Both neural network and local regression models are able to predict and reproduce the stage–discharge multiple loops that are observed at the outlet of a 28.5?km2 low-gradient subcatchment in southwestern Louisiana. However, the neural network model is characterized with higher prediction ability for most of the tested runoff events. In agreement with the physical characteristics of low-gradient streams, the results indicate the importance of including information about downstream and upstream water levels, in addition to water level at the prediction site.     

Modeling multivariate discrete failure time data

A bivariate discrete survival distribution that allows flexible modeling of the marginal distributions and yields a constant odds ratio at any grid point is proposed. The distribution can be extended to a multivariate distribution and is readily generalized to accommodate covariates in the marginal distributions and pairwise odds ratios. In addition, a pseudo-likelihood estimation procedure for estimating the regression coefficients in the marginal models and the association parameters in the pairwise odds ratios is presented. We evaluate the performance of the proposed estimation procedure through simulations. For bivariate data, pseudo-likelihood estimation of the association parameter has high efficiency. Loss of efficiency in the marginal regression coefficient estimates is small when the association is not strong. For both the marginal regression coefficients and the association parameter, coverage probabilities are close to the 95% nominal level. For multivariate data, the simulation results show that the parameter estimates are consistent. Coverage probability for the regression coefficient in the marginal model is close to the 95% nominal level but is slightly less than the nominal level for the association parameter. We illustrate the proposed methods using a subset of the Framingham Heart Study data where a significant positive association was found between the failure times of siblings.     

Simulated Micromechanical Models Using Artificial Neural Networks

A new method, termed simulated micromechanical models using artificial neural networks (MMANN), is proposed to generate micromechanical material models for nonlinear and damage behavior of heterogeneous materials. Artificial neural networks (ANN) are trained with results from detailed nonlinear finite-element (FE) analyses of a repeating unit cell (UC), with and without induced damage, e.g., voids or cracks between the fiber and matrix phases. The FE simulations are used to form the effective stress-strain response for a unit cell with different geometry and damage parameters. The FE analyses are performed for a relatively small number of applied strain paths and damage parameters. It is shown that MMANN material models of this type exhibit many interesting features, including different tension and compression response, that are usually difficult to model by conventional micromechanical approaches. MMANN material models can be easily applied in a displacement-based FE for nonlinear analysis of composite structures. Application examples are shown where micromodels are generated to represent the homogenized nonlinear multiaxial response of a unidirectional composite with and without damage. In the case of analysis with damage growth, thermodynamics with irreversible processes (TIP) is used to derive the response of an equivalent homogenized damage medium with evolution equations for damage. The proposed damage formulation incorporates the generalizations generated by the MMANN method for stresses and other possible responses from analysis results of unit cells with fixed levels of damage.     

Study of Properties of Intermetallic Compounds of Rare Earth Metals by Artificial Neural Networks

ＹａｎＬｕ－Ｍｉｎｇ；（严六明）；ＺｈａｎＱｉａｎ－Ｂａｏ；（詹千宝）；ＱｉｎＰｅｉ；（钦佩）；ＣｈｅｎＮｉａｎ－Ｙｉ；（陈念贻）（ＳｈａｎｇｈａｉＩｎｓｔｉｔｕｔｅｏｆＭｅｔａｌｌｕｒｇｙ，ＡｃａｄｅｍｉａＳｉｎｉｃａ，Ｓｈａｎｇｈａｉ２０００５０...     

Optimizing the Prediction Accuracy of Concrete Compressive Strength Based on a Comparison of Data-Mining Techniques

This study attempts to optimize the prediction accuracy of the compressive strength of high-performance concrete (HPC) by comparing data-mining methods. Modeling the dynamics of HPC, which is a highly complex composite material, is extremely challenging. Concrete compressive strength is also a highly nonlinear function of ingredients. Several studies have independently shown that concrete strength is determined not only by the water-to-cement ratio but also by additive materials. The compressive strength of HPC is a function of all concrete content, including cement, fly ash, blast-furnace slag, water, superplasticizer, age, and coarse and fine aggregate. The quantitative analyses in this study were performed by using five different data-mining methods: two machine learning models (artificial neural networks and support vector machines), one statistical model (multiple regression), and two metaclassifier models (multiple additive regression trees and bagging regression trees). The methods were developed and tested against a data set derived from 17 concrete strength test laboratories. The cross-validation of unbiased estimates of the prediction models for performance comparison purposes indicated that multiple additive regression tree (MART) was superior in prediction accuracy, training time, and aversion to overfitting. Analytical results suggested that MART-based modeling is effective for predicting the compressive strength of varying HPC age.     

Parameter Estimation in Groundwater Hydrology Using Artificial Neural Networks

The capability of artificial neural networks to act as universal function approximators has been traditionally used to model problems in which the relation between dependent and independent variables is poorly understood. In this paper, the capability of an artificial neural network to provide a data-driven approximation of the explicit relation between transmissivity and hydraulic head as described by the groundwater flow equation is demonstrated. Techniques are applied to determine the optimal number of nodes and training patterns needed for a neural network to approximate groundwater parameters for a simulated groundwater modeling case study. Furthermore, the paper explains how such an approximation can be used for the purpose of parameter estimation in groundwater hydrology.     

Replacing Outliers and Missing Values from Activated Sludge Data Using Kohonen Self-Organizing Map

Modeling the activated sludge wastewater treatment plant plays an important role in improving its performance. However, there are many limitations of the available data for model identification, calibration, and verification, such as the presence of missing values and outliers. Because available data are generally short, these gaps and outliers in data cannot be discarded but must be replaced by more reasonable estimates. The aim of this study is to use the Kohonen self-organizing map (KSOM), unsupervised neural networks, to predict the missing values and replace outliers in time series data for an activated sludge wastewater treatment plant in Edinburgh, U.K. The method is simple, computationally efficient and highly accurate. The results demonstrated that the KSOM is an excellent tool for replacing outliers and missing values from a high-dimensional data set. A comparison of the KSOM with multiple regression analysis and back-propagation artificial neural networks showed that the KSOM is superior in performance to either of the two latter approaches.     

Comparative study of the stress flow models for high-boron corrosion-resistant steel based on an Arrhenius-type equation and artificial neural networks

A comparative analysis is performed for the stress flow models of a steel containing 2.3 wt % B during its compressive hot deformation based on an Arrhenius-type equation and artificial neural networks. It is shown that the accuracy of the obtained data using the model based on artificial neural networks is much higher (relative calculation error is 0.2%) as compared to that of the regression model (relative error is 11%).     

Artificial Intelligence-Based Inductive Models for Prediction and Classification of Fecal Coliform in Surface Waters

This paper describes the use of inductive models developed using two artificial intelligence (AI)-based techniques for fecal coliform prediction and classification in surface waters. The two AI techniques used include artificial neural networks (ANNs) and a fixed functional set genetic algorithm (FFSGA) approach for function approximation. While ANNs have previously been used successfully for modeling water quality constituents, FFSGA is a relatively new technique of inductive model development. This paper will evaluate the efficacy of this technique for modeling indicator organism concentrations. In scenarios where process-based models cannot be developed and/or are not feasible, efficient and effective inductive models may be more suitable to provide quick and reasonably accurate predictions of indicator organism concentrations and associated water quality violations. The relative performance of AI-based inductive models is compared with conventional regression models. When raw data are used in the development of the inductive models described in this paper, the AI models slightly outperform the traditional regression models. However, when log transformed data are used, all inductive models show comparable performance. While the work validates the strength of simple regression models, it also validated FFSGA to be an effective technique that competes well with other state-of-the-art and complex techniques such as ANNs. FFSGA comes with the added advantage of resulting in a simple, easy to use, and compact functional form of the model sought. This work adds to the limited amount of research on the use of data-driven modeling methods for indicator organisms.     

Predicting the Endpoint Phosphorus Content of Molten Steel in BOF by Two-stage Hybrid Method

A two-stage hybrid method is proposed to predict the phosphorus content of molten steel at the endpoint of steelmaking in BOF (Basic Oxygen Furnace). At the first clustering stage, the weighted K-means is performed to produce clusters with homogeneous data. At the second predicting stage, each fuzzy neural network is carried out on each cluster and the results from all fuzzy neural networks are combined to be the final result of the hybrid method. The hybrid method and single fuzzy neural network are compared and the results show that the hybrid method outperforms single fuzzy neural network.     

Kinetic parameter estimation by numerical algorithms and multiple linear regression: theoretical

A new method is presented for the determination of kinetic parameters based on a functional relationship among experimental data derived from the postulated model. The data, even though containing errors, are manifestations of this relationship, which should be satisfied by parameters fitted to the system. The procedure involves the use of numerical integration and/or differentiation of the data, followed by multiple linear regression. It does not require initial estimates or repetitive iteration for linear systems and can be applied to nonlinear models. The accuracy of estimated parameter depends on the goodness of the particular numerical approximation method used.