• Journal of Information Processing Systems
• /
• 제20권4호
• /
• pp.501-513
• /
• 2024

In recent years, graph neural networks (GNNs) have been extensively used to analyze graph data across various domains because of their powerful capabilities in learning complex graph-structured data. However, recent research has focused on improving the performance of a single GNN with only two or three layers. This is because stacking layers deeply causes the over-smoothing problem of GNNs, which degrades the performance of GNNs significantly. On the other hand, ensemble methods combine individual weak models to obtain better generalization performance. Among them, gradient boosting is a powerful supervised learning algorithm that adds new weak models in the direction of reducing the errors of the previously created weak models. After repeating this process, gradient boosting combines the weak models to produce a strong model with better performance. Until now, most studies on GNNs have focused on improving the performance of a single GNN. In contrast, improving the performance of GNNs using multiple GNNs has not been studied much yet. In this paper, we propose gradient boosted graph neural networks (GBGNN) that combine multiple shallow GNNs with gradient boosting. We use shallow GNNs as weak models and create new weak models using the proposed gradient boosting-based loss function. Our empirical evaluations on three real-world datasets demonstrate that GBGNN performs much better than a single GNN. Specifically, in our experiments using graph convolutional network (GCN) and graph attention network (GAT) as weak models on the Cora dataset, GBGNN achieves performance improvements of 12.3%p and 6.1%p in node classification accuracy compared to a single GCN and a single GAT, respectively.

• Journal of Mechanical Science and Technology
• /
• 제15권5호
• /
• pp.580-591
• /
• 2001

This paper proposes a fuzzy-sliding mode control which is designed by a self tuning fuzzy inference method based on a genetic algorithm. Using the method, the number of inference rules and the shape of the membership functions of the proposed fuzzy-sliding mode control are optimized without the aid of an expert in robotics. The fuzzy outputs of the consequent part are updated by the gradient descent method. It is further guaranteed that the selected solution becomes the global optimal solution by optimizing Akaikes information criterion expressing the quality of the inference rules. In order to evaluate the learning performance of the proposed fuzzy-sliding mode control based on a genetic algorithm, a trajectory tracking simulation of the polishing robot is carried out. Simulation results show that the optimal fuzzy inference rules are automatically selected by the genetic algorithm and the trajectory control result is similar to the result of the fuzzy-sliding mode control which is selected through trial error by an expert. Therefore, a designer who does not have expert knowledge of robot systems can design the fuzzy-sliding mode controller using the proposed self tuning fuzzy inference method based on the genetic algorithm.

• The Journal of Korean Institute of Electromagnetic Engineering and Science
• /
• 제27권8호
• /
• pp.676-683
• /
• 2016

In this paper, an adaptive nulling algorithm which can be used to form nulls in the direction of jammer or interference signals in array antennas of single port system is proposed. The proposed adaptive algorithm does not require a priori knowledge of the incoming signal direction and can be applied to the partially adaptive arrays. This algorithm is the combination of the PSO(Particle Swam Optimization) algorithm and the gradient-based perturbation adaptive algorithm, which shows stable nulling performance adaptively even on the moving jammer environment where the incident direction of the interference signal is changing with time.

• Smart Media Journal
• /
• 제3권2호
• /
• pp.9-13
• /
• 2014

The Levenberg-Marquardt (LM) algorithm is the most widely used fitting algorithm. It outperforms simple gradient descent and other conjugate gradient methods in a wide variety of problems. Based on the work of paper[1], we propose a modified Levenberg-Marquardt algorithm for better performance of mobile system. The LM parameter at the $k_{th}$ iteration is chosen ${\mu}=A{\bullet}{\parallel}f(x){\parallel}{\bullet}I$ where f is the residual function, and J is the Jacobi of f. In this paper, we show this method is more efficient than traditional method under the situation that the system iteration is limited.

• The Journal of the Acoustical Society of Korea
• /
• 제38권3호
• /
• pp.283-289
• /
• 2019

RLS (Recursive-least-squares) algorithm is known to have good convergence and excellent error level after convergence. However, there is a disadvantage that numerical instability is included in the algorithm due to inverse matrix calculation. In this paper, we propose an algorithm with no matrix inversion to avoid the instability aforementioned. The proposed algorithm still keeps the same convergence performance. In the proposed algorithm, we adopt an averaged gradient-based step size as a self-adjusted step size. In addition, a variable forgetting factor is introduced to provide superior performance for time-varying channel estimation. Through simulations, we compare performance with conventional RLS and show its equivalency. It also shows the merit of the variable forgetting factor in time-varying channels.

• Journal of Korean Association for Spatial Structures
• /
• 제21권2호
• /
• pp.41-48
• /
• 2021

A smart tuned mass damper (TMD) is widely studied for seismic response reduction of various structures. Control algorithm is the most important factor for control performance of a smart TMD. This study used a Deep Deterministic Policy Gradient (DDPG) among reinforcement learning techniques to develop a control algorithm for a smart TMD. A magnetorheological (MR) damper was used to make the smart TMD. A single mass model with the smart TMD was employed to make a reinforcement learning environment. Time history analysis simulations of the example structure subject to artificial seismic load were performed in the reinforcement learning process. Critic of policy network and actor of value network for DDPG agent were constructed. The action of DDPG agent was selected as the command voltage sent to the MR damper. Reward for the DDPG action was calculated by using displacement and velocity responses of the main mass. Groundhook control algorithm was used as a comparative control algorithm. After 10,000 episode training of the DDPG agent model with proper hyper-parameters, the semi-active control algorithm for control of seismic responses of the example structure with the smart TMD was developed. The simulation results presented that the developed DDPG model can provide effective control algorithms for smart TMD for reduction of seismic responses.

• Journal of the Korean Society for Industrial and Applied Mathematics
• /
• 제22권1호
• /
• pp.1-13
• /
• 2018

Boosting, one of the most successful algorithms for supervised learning, searches the most accurate weighted sum of weak classifiers. The search corresponds to a convex programming with non-negativity and affine constraint. In this article, we propose a novel Conjugate Gradient algorithm with the Modified Polak-Ribiera-Polyak conjugate direction. The convergence of the algorithm is proved and we report its successful applications to boosting.

• 전자공학회논문지 IE
• /
• 제47권2호
• /
• pp.27-31
• /
• 2010

In this paper, a predictor switching algorithm for lossless compression is proposed. It uses adaptively one of two predictors using errors obtained by MED(median edge detector) and GAP(gradient adaptive prediction). The reduced error is measured by existing entropy method. Experimental results show that the proposed algorithm can compress higher than existing predictive methods.

• Phonetics and Speech Sciences
• /
• 제7권4호
• /
• pp.17-25
• /
• 2015

For blind source separation of convolutive mixtures, FDICA(Frequency Domain Independent Component Analysis) algorithms are generally used. Since FDICA algorithm such as Sawada FDICA, IVA(Independent Vector Analysis) works on the frequency bin basis with a natural gradient descent method, it takes much time to converge. In this paper, we propose a new method to improve convergence speed in FDICA algorithm. The proposed method reduces the number of iteration drastically in the process of natural gradient descent method by applying a weighted inner product constraint of unmixing matrix. Experimental results have shown that the proposed method achieved large improvement of convergence speed without degrading the separation performance of the baseline algorithms.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• 제29권8호
• /
• pp.903-910
• /
• 2005

Inverse radiation problems are solved for estimating the boundary conditions such as temperature distribution and wall emissivity in axisymmetric absorbing, emitting and scattering medium, given the measured incident radiative heat fluxes. Various regularization methods, such as hybrid genetic algorithm, conjugate-gradient method and finite-difference Newton method, were adopted to solve the inverse problem, while discussing their features in terms of estimation accuracy and computational efficiency. Additionally, we propose a new combined approach that adopts the hybrid genetic algorithm as an initial value selector and uses the finite-difference Newton method as an optimization procedure.