I-Cheng Yeh

Modeling of strength of high-performance concrete using artificial neural networks

Several studies independently have shown that concrete strength development is determined not only by the water-to-cement ratio, but that it also is influenced by the content of other concrete ingredients. High-performance concrete is a highly complex material, which makes modeling its behavior a very difficult task. This paper is aimed at demonstrating the possibilities of adapting artificial neural networks (ANN) to predict the compressive strength of high-performance concrete. A set of trial batches of HPC was produced in the laboratory and demonstrated satisfactory experimental results. This study led to the following conclusions: 1) A strength model based on ANN is more accurate than a model based on regression analysis; and 2) It is convenient and easy to use ANN models for numerical experiments to review the effects of the proportions of each variable on the concrete mix.

Knowledge discovery of concrete material using Genetic Operation Trees

This study proposed a novel knowledge discovery method, Genetic Operation Tree (GOT), which is composed of operation tree (OT) and genetic algorithm (GA), to automatically produce self-organized formulas to predict compressive strength of High-Performance Concrete. In GOT, OT plays the architecture to represent an explicit formula, and GA plays the optimization mechanism to optimize the OT to fit experimental data. Experimental data from several different sources were used to evaluate the method. The results showed that GOT can produce formulas which are more accurate than nonlinear regression formulas but less accurate than neural network models. However, neural networks are black box models, while GOT can produce explicit formulas, which is an important advantage in practical applications.

Hybrid genetic algorithms for optimization of truss structures

The genetic algorithm (GA) is a new optimization paradigm that models a natural evolution mechanism. The framework of the GA naturally corresponds to a discrete optimization problem. Although the GA is very robust, it is also very computationally intensive and hence slower than other methods. To speed up the convergence, this article proposes a hybrid GA that combines the concept of survival of the fittest with the concept of adaptation. The fully stressed design optimality criterion is employed to play the role of adaptation. Numerical examples show that even though the displacement constraints are active, (1) both average weight and minimum weight obtained by a hybrid GA are less than those obtained by a pure GA, (2) a hybrid GA is more stable than a pure GA, and (3) the speed of convergence of a hybrid GA is superior to that of a pure GA.

Application of neural networks to automatic soil pressure balance control for shield tunneling

Abstract In this paper, a shield control system software to balance soil pressure on the shield cutting face is described. This software, which adjusts the speed of the shield jack and the speed of the screw conveyor, is based on a neural network. The basic structure of the control system software consists of a modeling mechanism and a control mechanism. The modeling mechanism of this system has a learning function based on a back-propagation neural network to model the mechanism of the soil pressure in the soil room of the shield. The learning function renews the model in accordance with the historical records of shield operation. The control mechanism of this system has a searching function to find the optimal value of the desired speed of the shield jack and the screw conveyor to reach the desired soil pressure. This system has been tested on a tunneling project in Taipei City. The results showed that the control method of this system is very effective as a means of controlling the shield in various start states.

Modeling chaotic two-dimensional mapping with fuzzy-neuron networks

Abstract The neural networks performance can be measured by efficiency and accuracy. The major disadvantages of neural network approach are that the generalization capability of neural networks is often significantly low, and it may take a very long time to tune the weights in the net to generate an accurate model for a highly complex and nonlinear systems. This paper presents a novel neural network architecture based on fuzzy set theory, fuzzy-neuron network (FNN), and tests its efficiency and accuracy in modeling chaotic two-dimensional mapping. The FNN architecture is a standard back-propagation neural network; however, fuzzy neurons are added to the network. The network is tested using two classic chaotic dynamic systems: Ikeda mapping and Jin mapping. Experimental results demonstrate that the fuzzy neurons in the network provide an enhanced network architecture and improve the performance of these networks significantly.

Evaluation approach to stock trading system using evolutionary computation

The past researches emphasize merely the avoidance of over-learning at the system level and ignore the problem of over-learning at the model level, which lead to the poor performance of the evolutionary computation based stock trading decision-making system. This study presents a new evaluation approach to focus on evaluating the generalization capability at the model level. An empirical study was provided and the results reveal four important findings. First, the decision-making system generated at the model design stage outperforms the system generated at the model validation stage, which shows over-learning at the model level. Secondly, for the decision-making system generated either at the model design stage or at the model validation stage, the investment performance in the training period is much better than that in the testing period, exhibiting over-learning at the system level. Third, employing moving timeframe approach is unable to improve the investment performance at the model validation stage. Fourth, reducing the evolution generation and input variables are unable to avoid the over-learning at the model level. The major contribution of this study is to clarify the issue of over-learning at the model and the system level. For future research, this study developed a more reliable evaluation approach in examining the generalization capability of evolutionary computation based decision-making system.

Cosmetics purchasing behavior - An analysis using association reasoning neural networks

The research objective is to propose a novel association analysis approach using association reasoning neural networks (ARNN) to discover the association rules from cosmetics purchasing. ARNN is modified from multi-layered perceptron and back-propagation algorithm. The number of association rules is controlled by the rule threshold and the number of hidden units. To explore the possibility of producing useful and meaningful association rules using ARNN, our study uses the practical cosmetics transaction data. The results show (1) the predicted output values of ARNN are close to their desired confidence values, (2) reducing the number of hidden units of ARNN can inhibit the generation of association rules with low support, and (3) ARNN has the ability of discovering the cohesion and expansion commodities and this information could be used to make pricing strategy. Therefore, ARNN could be a promising alternative approach for association analysis.

Structural Engineering Applications with Augmented Neural Networks

This paper presents an augmented neural network (ANN), a novel neural network architecture, and examines its efficiency and accuracy for structural engineering applications. The proposed architecture is that of a standard backpropagation neural network with augmented neurons, that is, logarithm neurons and exponent neurons are added to the networks input and output layers. The principles of augmented neural networks are (1) the augmented neurons are highly sensitive in the boundary domain, thereby facilitating construction of accurate mapping in the models boundary domain, and (2) the network denotes each input variable with multiple input neurons, thus allowing a highly interactive function on hidden neurons to be easily formed. Therefore, the hidden neurons can more easily construct an accurate network output for a highly interactive mapping model. Experimental results demonstrate that the networks logarithm and exponent neurons provide a markedly enhanced network architecture capable of improving the networks performance for structural engineering applications.

Supervised Learning Probabilistic Neural Networks

This study proposed supervised learning probabilistic neural networks (SLPNN) which have three kinds of network parameters: variable weights representing the importance of input variables, the reciprocal of kernel radius representing the effective range of data, and data weights representing the data reliability. These three kinds of parameters can be adjusted through training. We tested three artificial functions as well as 15 benchmark problems, and compared it with multi-layered perceptron (MLP) and probabilistic neural networks (PNN). The results showed that SLPNN is slightly more accurate than MLP, and much more accurate than PNN. Besides, the data weights can find the noise data in data set, and the variable weights can measure the importance of input variables and have the greatest contribution to accuracy of model among the three kinds of network parameters.

CASE-BASED APPROACHES FOR PRELIMINARY DESIGN OF STEEL BUILDING FRAMES

Case-based reasoning (CBR) is a technique for solving new problems by adapting solutions that were obtained by solving old problems. In this article, three distinct CBR methodologies are examined for their efficiency and accuracy in modeling steel building frame design, including conventional nearest neighbor CBR and two novel CBR methodologies, collaborative CBR, which combines several cases to generate a better solution, and hybrid CBR, which hybridizes two important CBR mechanisms, adaptation and combination, to obtain a more accurate solution. The results show that the hybrid CBR techniques are slightly more accurate than back-propagation neural networks and much more accurate than the other two CBR methodologies.