Committee Of Neural Networks Research Articles

An Organized Committee of Artificial Neural Networks in the Classification of Human Chromosomes

networks are organized in committees to improve the correctness of the decisions created by artificial neural networks (ANN's). In the classification of human chromosomes, it is accustomed to use multilayer perceptrons with multiple (22-24) outputs. Because of the huge number of synaptic weights to be tuned, these classifiers cannot go beyond a level of 92% overall correctness. In this study we represent a special organized committee of 462 simple perceptrons to improve the rate of correct classification of 22 types of human chromosomes. Each of these simple perceptrons is trained to distinguish between two types of chromosomes. When a new data is entered, the votes of these 462 simple perceptrons and additional 22 dummy perceptrons create a decision matrix of the size 22×22. By a special assembling of these votes we get a higher rate of correct classification of 22 types of human chromosomes.

군집 신경망기법을 이용한 해상풍력발전기 지지구조물의 건전성 모니터링 기법

A damage estimation method for monopile support structure of offshore wind turbine using modal properties and committee of neural networks is presented for effective structural health monitoring. An analytical model for a monopile support structure is established, and the natural frequencies, mode shapes, and mode shape slopes for the support structure are calculated considering soil condition and added mass. The input to the neural networks consists of the modal properties and the output is composed of the stiffness indices of the support structure. Multiple neural networks are constructed and each individual network is trained independently with different initial synaptic weights. Then, the estimated stiffness indices from different neural networks are averaged. Ten damage cases are estimated using the proposed method, and the identified damage locations and severities agree reasonably well with the exact values. The accuracy of the estimation can be improved by applying the committee of neural networks which is a statistical approach averaging the damage indices in the functional space.

A neural-network committee machine approach to the inverse kinematics problem solution of robotic manipulators

In robotics, inverse kinematics problem solution is a fundamental problem in robotics. Many traditional inverse kinematics problem solutions, such as the geometric, iterative, and algebraic approaches, are inadequate for redundant robots. Recently, much attention has been focused on a neural-network-based inverse kinematics problem solution in robotics. However, the result obtained from the neural network requires to be improved for some sensitive tasks. In this paper, a neural-network committee machine (NNCM) was designed to solve the inverse kinematics of a 6-DOF redundant robotic manipulator to improve the precision of the solution. Ten neural networks (NN) were designed to obtain a committee machine to solve the inverse kinematics problem using separately prepared data set since a neural network can give better result than other ones. The data sets for the neural-network training were prepared using prepared simulation software including robot kinematics model. The solution of each neural network was evaluated using direct kinematics equation of the robot to select the best one. As a result, the committee machine implementation increased the performance of the learning.

The Impact of Macroeconomic Factors on Residential Property Price Indices in Europe

Abstract This paper aims to determine the influence of selected variables on residential property price indices for the European countries, with particular attention paid to Italy and Poland, using a rough set theory and an approach that uses a committee of artificial neural networks. Additionally, the overall analysis for each European country is presented. Quarterly time series data constituted the material for testing and empirical results. The developed models show that the economic and financial situation of European countries affects residential property markets. Residential property markets are connected, despite the fact that they are situated in different parts of Europe. The economic and financial crisis of countries has variable influence on prices of real estate. The results also suggest that methodology based on the rough set theory and a committee of artificial neural networks has the ability to learn, generalize, and converge the residential property prices index.

Quantile based decision making rule of the neural networks committee for ill-posed approximation problems

In this paper a multidimensional function approximation problem is stated. This problem is characterized by the strong influence of arguments measurement errors on the accuracy of the function estimation and a small amount of train data. A neural networks based solution is used for this problem.To improve the accuracy of the approximation model it is proposed to use a neural networks committee with an original decision making rule for the construction of the generalized function estimate. The developed rule is based on a specially introduced indirect accuracy measure and α-quantile calculation of its probability level. The decision making rule is trained on a set of patterns and uses statistical properties of each pattern's processing by the committee's networks.The computational scheme of the proposed approximation model and the effectiveness of the proposed approach are demonstrated on a simple model example. The developed method was successfully applied to a real industrial problem of metal's hardness characteristics estimation on the basis of kinetic indentation data. The results of modeling experiments are discussed.

Detecting the Authors of Texts by Neural Network Committee Machines

This paper proposes a means of using a boosting by filtering algorithm in artificial neural networks to identify the author of a text. This approach involves filtering the training examples by different versions of a weak learning algorithm. It assures the availability of a large source of examples, with the examples being either discarded or kept during training. An advantage of this approach is that it allows for a small memory requirement. Once the network has been trained, its hidden layer activations are recorded as a representation of the selected lexical descriptors of an author. This stored information can then be used to identify the texts written by the same author. Texts studied are literary works of two Bosnian writers, Ivo Andrić (1892-1975) and M. Meša Selimović (1910-1982). The data collected by counting syntactic characteristics in 1466 paragraphs of "na drini ćupria" by Ivo Andrić, and "derviš i smirt" by M. Meša Selimović each.

The impact of macroeconomic factors on residential property prices indices in Europe

This paper aims to determine the influence of selected variables on residential property price indices for European countries, with particular attention paid to Italy and Poland, using a rough set theory and an approach that uses a committee of artificial neural networks. Additionally, the overall analysis for each European country is presented. Quarterly time series data constituted the material for testing and empirical results. The developed models show that the economic and financial situation of European countries affects residential property markets. Residential property markets are connected, despite the fact that they are situated in different parts of Europe. The economic and financial crisis of countries has variable influence on prices of real estate. The results also suggest that methodology based on the rough set theory and a committee of artificial neural networks has the ability to learn, generalize, and converge the residential property prices index.

Bayesian neural networks for the prediction of stochastic travel times in urban networks

Urban travel time prediction has received much less attention than predictions on freeways, perhaps because urban travel times show much larger variations and are therefore much harder to predict. However, urban travel time can form a substantial part of the total travel time of a road user and therefore effort should be taken to predict urban travel times. In this study, neural networks are used for urban travel time prediction because these have shown to be able to deal with noisy data. Bayesian techniques are used for training of the networks, resulting in committees with lower error and in confidence bounds. It is shown that the neural network committees are capable of predicting the ‘low-frequency trend’, which can be seen when the high-frequency component of travel time is removed using de-noising. The errors of the predictions on the low-frequency trend are in the same order as when predicting freeway travel times, and it is shown that the predicted confidence bounds are accurate.

Neural networks committee decision making for estimation of metal’s hardness properties from indentation data

In this paper the problem of metal's hardness properties estimation from indentation data is concerned. This problem belongs to a class of ill-posed vector function approximation problems and can't be solved by a single multilayered perceptron at the required precision level. A special neural networks committee architecture is developed in order to obtain precise estimates of metal's hardness properties. This method involves well-posed direct indentation task solution and a quantile based idea for best estimates selection. Studies have shown the estimates produced by the committee to be stable even in the case of noise presence that is similar to the true experimental one.

Robust artificial intelligence tool for automatic start-up of the supplementary medium feeding in recombinant E. coli cultivations

One of the most important events in fed-batch fermentations is the definition of the moment to start the feeding. This paper presents a methodology for a rational selection of the architecture of an artificial intelligence (AI) system, based on a neural network committee (NNC), which identifies the end of the batch phase. The AI system was successfully used during high cell density cultivations of recombinant Escherichia coli. The AI algorithm was validated for different systems, expressing three antigens to be used in human and animal vaccines: fragments of surface proteins of Streptococcus pneumoniae (PspA), clades 1 and 3, and of Erysipelothrix rhusiopathiae (SpaA). Standard feed-forward neural networks (NNs), with a single hidden layer, were the basis for the NNC. The NN architecture with best performance had the following inputs: stirrer speed, inlet air, and oxygen flow rates, carbon dioxide evolution rate, and CO2 molar fraction in the exhaust gas.

Determination of anthocyanin concentration in whole grape skins using hyperspectral imaging and adaptive boosting neural networks

This paper reports a novel application of a type of neural network committee, called AdaBoost, to the estimation of grape anthocyanin concentration using hyperspectral data. The inputs from the neural networks were the principal components of the grapes’ spectra. Hyperspectral data were collected in the reflectance mode for 46 individual whole grapes of the Cabernet Sauvignon variety, using a hyperspectral camera that operates with wavelengths ranging from 400 to 1000 nm at an approximate 0.6 nm resolution. The hyperspectral camera was positioned a few tens of centimetres away from the grapes. The grapes were harvested on five dates between August 28th and September 23rd in 2009 and presented average sugar content values between 14.6 and 20.2 Brix. They were kept frozen until January 2010, when they were thawed and the hyperspectral data collected at ambient temperature. The anthocyanin concentration values obtained by our calibrations exhibited a squared correlation coefficient value of 0.65 compared to the values measured using conventional laboratory techniques. This correlation value is better than the value reported in another recent scientific work which estimated anthocyanin values in individual whole grapes of Cabernet Sauvignon.

NFL Prediction using Committees of Artificial Neural Networks

This paper analyzes the ability of a neural network model to predict the outcome of NFL games. This model uses only readily available statistics, such as passing yards, rushing yards, fumbles lost, and scoring. A key component of this model is the use of statistical differentials to compare teams. For example, the offensive passing yards gained by one team are compared to the defensive passing yards allowed by an opposing team to create a data set of expected values for a given matchup. By using principal component analysis and derivative based analysis, we determined which statistics influence our model the most. We assessed the performance of the model by comparing its performance to that of published prediction algorithms and the Las Vegas oddsmakers over multiple seasons. Two novel aspects of this work include the use of multiple committees of machines for prediction and the use of our model to simulate virtual round-robin tournaments to establish an objective ranking of the teams.

Molecular Dynamics Investigation of the Bimolecular Reaction BeH + H2 → BeH2 + H on an ab Initio Potential-Energy Surface Obtained Using Neural Network Methods with Both Potential and Gradient Accuracy Determination

The classical reaction dynamics of a four-body, bimolecular reaction on a neural network (NN) potential-energy surface (PES) fitted to a database obtained solely from ab initio MP2/6-311G(d,p) calculations are reported. The present work represents the first reported application of ab initio NN methods to a four-body, bimolecular, gas-phase reaction where bond extensions reach 8.1 A for the BeH + H(2) --> BeH(2) + H reaction. A modified, iterative novelty sampling method is used to select data points based on classical trajectories computed on temporary NN surfaces. After seven iterations, the sampling process is found to converge after selecting 9604 configurations. Incorporation of symmetry increases this to 19 208 BeH(3) configurations. The analytic PES for the system is obtained from the ensemble average of a five-member (6-60-1) NN committee. The mean absolute error (MAE) for the committee is 0.0046 eV (0.44 kJ mol(-1)). The total energy range of the BeH(3) database is 147.0 kJ mol(-1). Therefore, this MAE represents a percent energy error of 0.30%. Since it is the gradient of the PES that constitutes the most important quantity in molecular dynamics simulations, the paper also reports mean absolute error for the gradient. This result is 0.026 eV A(-1) (2.51 kJ mol(-1) A(-1)). Since the gradient magnitudes span a range of 15.32 eV A(-1) over the configuration space tested, this mean absolute gradient error represents a percent error of 0.17%. The mean percent absolute relative gradient error is 4.67%. The classically computed reaction cross sections generally increase with total energy. They vary from 0.007 to 0.030 A(2) when H(2) is at ground state, and from 0.05 to 0.10 A(2) when H(2) is in the first excited state. Trajectory integration is very fast using the five-member NN PES. The average trajectory integration time is 1.07 s on a CPU with a clock speed of 2.4 GHz. Zero angular momentum collisions are also investigated and compared with previously reported quantum dynamics on the same system. The quantum reaction probabilities exhibit pronounced resonance effects that are absent in the classical calculations. The magnitudes of quantum and classical results are in fair accord with the classical results being about 30-40% higher due to the lack of quantum restrictions on the zero-point vibrational energy.

Bayesian committee of neural networks to predict travel times with confidence intervals

Short-term prediction of travel time is one of the central topics in current transportation research and practice. Among the more successful travel time prediction approaches are neural networks and combined prediction models (a ‘committee’). However, both approaches have disadvantages. Usually many candidate neural networks are trained and the best performing one is selected. However, it is difficult and arbitrary to select the optimal network. In committee approaches a principled and mathematically sound framework to combine travel time predictions is lacking. This paper overcomes the drawbacks of both approaches by combining neural networks in a committee using Bayesian inference theory. An ‘evidence’ factor can be calculated for each model, which can be used as a stopping criterion during training, and as a tool to select and combine different neural networks. Along with higher prediction accuracy, this approach allows for accurate estimation of confidence intervals for the predictions. When comparing the committee predictions to single neural network predictions on the A12 motorway in the Netherlands it is concluded that the approach indeed leads to improved travel time prediction accuracy.

Neural network committees for finger joint angle estimation from surface EMG signals

BackgroundIn virtual reality (VR) systems, the user's finger and hand positions are sensed and used to control the virtual environments. Direct biocontrol of VR environments using surface electromyography (SEMG) signals may be more synergistic and unconstraining to the user. The purpose of the present investigation was to develop a technique to predict the finger joint angle from the surface EMG measurements of the extensor muscle using neural network models.MethodologySEMG together with the actual joint angle measurements were obtained while the subject was performing flexion-extension rotation of the index finger at three speeds. Several neural networks were trained to predict the joint angle from the parameters extracted from the SEMG signals. The best networks were selected to form six committees. The neural network committees were evaluated using data from new subjects.ResultsThere was hysteresis in the measured SMEG signals during the flexion-extension cycle. However, neural network committees were able to predict the joint angle with reasonable accuracy. RMS errors ranged from 0.085 ± 0.036 for fast speed finger-extension to 0.147 ± 0.026 for slow speed finger extension, and from 0.098 ± 0.023 for the fast speed finger flexion to 0.163 ± 0.054 for slow speed finger flexion.ConclusionAlthough hysteresis was observed in the measured SEMG signals, the committees of neural networks were able to predict the finger joint angle from SEMG signals.

An algorithm for automated authorship attribution using neural networks

We present an algorithm as evidence of the possibility of a truly automated stylometric authorship attribution tool, based on committees of artificial neural networks. Neural networks have an advantage over traditional statistical stylometry in that they are inherently nonlinear, and therefore can consider nonlinear interactions between stylometric variables. The algorithm presented (1) is intended to demonstrate the feasibility of an automated approach using neural networks and (2) highlights important areas for further research. We present results of two separate test experiments-Shakespeare and Marlowe, and the Federalist Papers-as a demonstration of the method's generality. In both cases, our algorithm produces committees that correctly predict the test works, without requiring the usual precursory statistical study to determine efficacious stylometric measures.

Ensemble of metamodels with optimized weight factors

Approximate mathematical models (metamodels) are often used as surrogates for more computationally intensive simulations. The common practice is to construct multiple metamodels based on a common training data set, evaluate their accuracy, and then to use only a single model perceived as the best while discarding the rest. This practice has some shortcomings as it does not take full advantage of the resources devoted to constructing different metamodels, and it is based on the assumption that changes in the training data set will not jeopardize the accuracy of the selected model. It is possible to overcome these drawbacks and to improve the prediction accuracy of the surrogate model if the separate stand-alone metamodels are combined to form an ensemble. Motivated by previous research on committee of neural networks and ensemble of surrogate models, a technique for developing a more accurate ensemble of multiple metamodels is presented in this paper. Here, the selection of weight factors in the general weighted-sum formulation of an ensemble is treated as an optimization problem with the desired solution being one that minimizes a selected error metric. The proposed technique is evaluated by considering one industrial and four benchmark problems. The effect of different metrics for estimating the prediction error at either the training data set or a few validation points is also explored. The results show that the optimized ensemble provides more accurate predictions than the stand-alone metamodels and for most problems even surpassing the previously reported ensemble approaches.

Neural network committee-based sensitivity analysis strategy for geotechnical engineering problems

Neural network usually acts as a “black box” in diverse fields to perform prediction, classification, and regression. Different from the conventional usages, neural network is herein attempted to handle factor sensitivity analysis in a geotechnical engineering system. After systematically investigating instability of employing single neural network in factor sensitivity analysis, a neural network committee (NNC)-based sensitivity analysis strategy is first algorithmically presented based on the particular mathematical ideas of weak law of large numbers in probability and optimization. Significantly, this study especially emphasizes the practical application of the NNC-based sensitivity analysis strategy to highlight the mechanism underlying in strata movement. The principal goal is to reveal the relationships among influential factors on strata movement through estimating the relative contribution of each explicative (input) variable on dependent (output) variables of strata movement. It is demonstrated that the NNC-based sensitivity analysis strategy rationally not only reveals the relative contribution of each explicative variable on dependent variables but also indicates the predictability of each dependent variable. In addition, an improved prediction model is resulted from integrating the sensitivity analysis results into neural network modeling, and it is capable of facilitating the convergence training of neural network model and advancing its prediction precision on strata movement angles. The above outcomes indicate that the NNC-based sensitivity analysis strategy provides a new paradigm of applying neural networks to deal with complex geotechnical engineering problems.

Online filtering of CO2 signals from a bioreactor gas outflow using a committee of constructive neural networks

This work proposes a committee of Cascade Correlation neural networks as an online smoother for random measurement noise. The goals of this paper are twofold: first it intends to explore the possibilities of using a constructive neural network algorithm to learn how to filter typical noisy data from a bioreactor, CO(2) mol fractions of the effluent gas during the aerobic cultivation of Bacillus megaterium to produce the enzyme penicillin G acylase. Second, to propose a committee of such networks for achieving more realistic results, capturing the inherent trend of the process. In order to do that this paper discusses the advantages of using a constructive neural network algorithm, describes how the committee of NNs operates and evaluates its performance using real CO(2) online data obtained in laboratorial experiments. The paper also presents results obtained with classical filtering algorithms, for comparison.

An incremental neural learning framework and its application to vehicle diagnostics

This paper presents a framework for incremental neural learning (INL) that allows a base neural learning system to incrementally learn new knowledge from only new data without forgetting the existing knowledge. Upon subsequent encounters of new data examples, INL utilizes prior knowledge to direct its incremental learning. A number of critical issues are addressed including when to make the system learn new knowledge, how to learn new knowledge without forgetting existing knowledge, how to perform inference using both the existing and the newly learnt knowledge, and how to detect and deal with aged learnt systems. To validate the proposed INL framework, we use backpropagation (BP) as a base learner and a multi-layer neural network as a base intelligent system. INL has several advantages over existing incremental algorithms: it can be applied to a broad range of neural network systems beyond the BP trained neural networks; it retains the existing neural network structures and weights even during incremental learning; the neural network committees generated by INL do not interact with one another and each sees the same inputs and error signals at the same time; this limited communication makes the INL architecture attractive for parallel implementation. We have applied INL to two vehicle fault diagnostics problems: end-of-line test in auto assembly plants and onboard vehicle misfire detection. These experimental results demonstrate that the INL framework has the capability to successfully perform incremental learning from unbalanced and noisy data. In order to show the general capabilities of INL, we also applied INL to three general machine learning benchmark data sets. The INL systems showed good generalization capabilities in comparison with other well known machine learning algorithms.