• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.157-165
• /
• 2020

In this paper. we propose an effective compensation scheme to the quantization error arisen from quantized learning in a machine learning on an embedded system. In the machine learning based on a gradient descent or nonlinear signal processing, the quantization error generates early vanishing of a gradient and occurs the degradation of learning performance. To compensate such quantization error, we derive an orthogonal compensation vector with respect to a maximum component of the gradient vector. Moreover, instead of the conventional constant learning rate, we propose the adaptive learning rate algorithm without any inner loop to select the step size, based on a nonlinear optimization technique. The simulation results show that the optimization solver based on the proposed quantized method represents sufficient learning performance.

• Geomechanics and Engineering
• /
• v.36 no.3
• /
• pp.259-276
• /
• 2024

The undrained shear strength is widely acknowledged as a fundamental mechanical property of soil and is considered a critical engineering parameter. In recent years, researchers have employed various methodologies to evaluate the shear strength of soil under undrained conditions. These methods encompass both numerical analyses and empirical techniques, such as the cone penetration test (CPT), to gain insights into the properties and behavior of soil. However, several of these methods rely on correlation assumptions, which can lead to inconsistent accuracy and precision. The study involved the development of innovative methods using extreme gradient boosting (XGB) to predict the pile set-up component "A" based on two distinct data sets. The first data set includes average modified cone point bearing capacity (q_t), average wall friction (f_s), and effective vertical stress (σ_vo), while the second data set comprises plasticity index (PI), soil undrained shear cohesion (S_u), and the over consolidation ratio (OCR). These data sets were utilized to develop XGBoost-based methods for predicting the pile set-up component "A". To optimize the internal hyperparameters of the XGBoost model, four optimization algorithms were employed: Particle Swarm Optimization (PSO), Social Spider Optimization (SSO), Arithmetic Optimization Algorithm (AOA), and Sine Cosine Optimization Algorithm (SCOA). The results from the first data set indicate that the XGBoost model optimized using the Arithmetic Optimization Algorithm (XGB - AOA) achieved the highest accuracy, with R2 values of 0.9962 for the training part and 0.9807 for the testing part. The performance of the developed models was further evaluated using the RMSE, MAE, and VAF indices. The results revealed that the XGBoost model optimized using XGBoost - AOA outperformed other models in terms of accuracy, with RMSE, MAE, and VAF values of 0.0078, 0.0015, and 99.6189 for the training part and 0.0141, 0.0112, and 98.0394 for the testing part, respectively. These findings suggest that XGBoost - AOA is the most accurate model for predicting the pile set-up component.

• Geophysics and Geophysical Exploration
• /
• v.10 no.4
• /
• pp.314-321
• /
• 2007

Recently, multi-dimensional joint inversion of geophysical data based on fundamentally different physical properties is being actively studied. Joint inversion can provide a way to obtaining much more accurate image of the subsurface structure. Through the joint inversion, furthermore, it is possible to directly estimate non-geophysical material properties from geophysical measurements. In this study, we developed a new algorithm for jointly inverting dc resistivity and seismic traveltime data based on the multiple constraints: (1) structural similarity based on cross-gradient, (2) correlation between two different material properties, and (3) a priori information on the material property distribution. Through the numerical experiments of surface dc resistivity and seismic refraction surveys, the performance of the proposed algorithm was demonstrated and the effects of different regularizations were analyzed. In particular, we showed that the hidden layer problem in the seismic refraction method due to an inter-bedded low velocity layer can be solved by the joint inversion when appropriate constraints are applied.

• Journal of Internet Computing and Services
• /
• v.15 no.3
• /
• pp.71-77
• /
• 2014

The equalization algorithm based on the cross-information potential concept and Dirac-delta functions (CIPD) has outstanding ISI elimination performance even under impulsive noise environments. The main drawback of the CIPD algorithm is a heavy computational burden caused by the use of a block processing method for its weight update process. In this paper, for the purpose of reducing the computational complexity, a new method of the gradient calculation is proposed that can replace the double summation with a single summation for the weight update of the CIPD algorithm. In the simulation results, the proposed method produces the same gradient learning curves as the CIPD algorithm. Even under strong impulsive noise, the proposed method yields the same results while having significantly reduced computational complexity regardless of the number of block data, to which that of the e conventional algorithm is proportional.

• Journal of IKEEE
• /
• v.23 no.4
• /
• pp.1302-1308
• /
• 2019

This paper presents an approach for classifying image made by acquired position data from a 2D Lidar sensor with a convolutional neural network (CNN). Lidar sensor has been widely used for unmanned devices owing to advantages in term of data accuracy, robustness against geometry distortion and light variations. A CNN algorithm consists of one or more convolutional and pooling layers and has shown a satisfactory performance for image classification. In this paper, different types of CNN architectures based on training methods, Gradient Descent(GD) and Levenberg-arquardt(LM), are implemented. The LM method has two types based on the frequency of approximating Hessian matrix, one of the factors to update training parameters. Simulation results of the LM algorithms show better classification performance of the image data than that of the GD algorithm. In addition, the LM algorithm with more frequent Hessian matrix approximation shows a smaller error than the other type of LM algorithm.

• Transactions on Control, Automation and Systems Engineering
• /
• v.2 no.2
• /
• pp.85-91
• /
• 2000

This paper proposes a novel method with pseudo-on-line scheme using the optimized look-up table based on the genetic algorithm which does not use the gradient and finds the global optimum of an un-constraint optimization problem. The technique is a pseudo-on-line method that optimally estimates the parameters of fuzzy PID(FPID) controller for systems with non-linearity, using the genetic algorithm. The proposed controller(GFPID) with the auto-tuning function is applied to the on-line and real-time control of speed at 3-phase induction motor, and its computer simulation is carried out. simulation results show that the proposed methodis more excellent that conventional FPID and PID controllers.

• Journal of the Korea Computer Graphics Society
• /
• v.23 no.3
• /
• pp.47-54
• /
• 2017

In this paper, we propose an improved curvature-based GPU (Graphics Processing Unit) isosurface ray-casting technique. Our method adopts the fast evaluation method proposed by Sigg et al. [1] to find the isosurface, but replaces the computation of the gradient and Hessian with the de Boor algorithm. In this way, we can reduce the number of additional texture fetches from 84 to 27 thus improving the performance by up to ${\approx}30%$, depending on the platforms.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.28 no.10C
• /
• pp.949-957
• /
• 2003

This paper presents a gradient ON-OFF algorithm of which the performance is very robust even when the angle spread increases in the mobile communication environments. The proposed method getting the diversity gain by utilizing the primary and secondary eigenvector, which corresponds to the largest and the second largest eigenvalue of the autocovariance matrix of the received signal vector, outperforms the method which just utilizes one eigenvector. By applying the proposed method to IS-2000 1X signal environments, it is observed that the proposed method shows excellent performance compared to a typical beamforming method using just one eigenvector, which considerably degrades the receiving performance as the angle spread increases.

• Journal of Sensor Science and Technology
• /
• v.25 no.6
• /
• pp.431-434
• /
• 2016

Currently, the radial basis function network (RBFN) and various other neural networks are employed to classify gases using chemical sensors arrays, and their performance is steadily improving. In particular, the identification performance of the RBFN algorithm is being improved by optimizing parameters such as the center, width, and weight, and improved algorithms such as the radial basis function network-stochastic gradient (RBFN-SG) and radial basis function network-normalized stochastic gradient (RBFN-NSG) have been announced. In this study, we optimized the number of centers, which is one of the parameters of the RBFN-NSG algorithm, and observed the change in the identification performance. For the experiment, repeated measurement data of 8 samples were used, and the elbow method was applied to determine the optimal number of centers for each sample of input data. The experiment was carried out in two cases(the only one center per sample and the optimal number of centers obtained by elbow method), and the experimental results were compared using the mean square error (MSE). From the results of the experiments, we observed that the case having an optimal number of centers, obtained using the elbow method, showed a better identification performance than that without any optimization.

• Steel and Composite Structures
• /
• v.44 no.5
• /
• pp.691-705
• /
• 2022

Shear failure in reinforced concrete (RC) structures is very hazardous. This failure is rarely predicted and may occur without any prior signs. Accurate shear strength prediction of the RC members is challenging, and traditional methods have difficulty solving it. This study develops a JAYA-GBRT model based on the JAYA algorithm and the gradient boosting regression tree (GBRT) to predict the shear strength of RC slender beams without stirrups. Firstly, 484 tests are carefully collected and divided into training and test sets. Then, the hyperparameters of the GBRT model are determined using the JAYA algorithm and 10-fold cross-validation. The performance of the JAYA-GBRT model is compared with five well-known empirical models. The comparative results show that the JAYA-GBRT model (R² = 0.982, RMSE = 9.466 kN, MAE = 6.299 kN, µ = 1.018, and Cov = 0.116) outperforms the other models. Moreover, the predictions of the JAYA-GBRT model are globally and locally explained using the Shapley Additive exPlanation (SHAP) method. The effective depth is determined as the most crucial parameter influencing the shear strength through the SHAP method. Finally, a Graphic User Interface (GUI) tool and a web application (WA) are developed to apply the JAYA-GBRT model for rapidly predicting the shear strength of RC slender beams without stirrups.