• Journal of Computing Science and Engineering
• /
• v.12 no.1
• /
• pp.24-35
• /
• 2018

Deep convolutional neural network (DCNN) is an advanced technology in image recognition. Because of extreme computing resource requirements, DCNN implementation with software alone cannot achieve real-time requirement. Therefore, the need to implement DCNN accelerator hardware is increasing. In this paper, we present a field programmable gate array (FPGA)-based hardware accelerator design of DCNN targeting handwritten Hangul character recognition application. Also, we present design optimization techniques in SDAccel environments for searching the optimal FPGA design space. The techniques we used include memory access optimization and computing unit parallelism, and data conversion. We achieved about 11.19 ms recognition time per character with Xilinx FPGA accelerator. Our design optimization was performed with Xilinx HLS and SDAccel environment targeting Kintex XCKU115 FPGA from Xilinx. Our design outperforms CPU in terms of energy efficiency (the number of samples per unit energy) by 5.88 times, and GPGPU in terms of energy efficiency by 5 times. We expect the research results will be an alternative to GPGPU solution for real-time applications, especially in data centers or server farms where energy consumption is a critical problem.

• Journal of Korean Society of Transportation
• /
• v.14 no.1
• /
• pp.51-67
• /
• 1996

The optimal path-finding problem becomes complicated when multiple variables are simultaneously considered such as physical route length, degree of congestion, traffic capacity of intersections, number of intersections and lanes, and existence of free ways. Therefore, many researchers in various fields (management science, computer science, applied mathematics, production planning, satellite launching) attempted to solve the problem by ignoring many variables for problem simplification, by developing intelligent algorithms, or by developing high-speed hardware. In this research, an integration of expert system technique and case-based reasoning in high level with a conventional algorithms in lower level was attempted to develop an optimal path-finding system. Early application of experienced driver's knowledge and case data accumulated in case base drastically reduces number of possible combinations of optimal paths by generating promising alternatives and by eliminating non-profitable alternatives. Then, employment of a conventional optimization algorithm provides faster search mechanisms than other methods such as bidirectional algorithm and $A^*$ algorithm. The conclusion obtained from repeated laboratory experiments with real traffic data in Seoul metropolitan area shows that the integrated approach to finding optimal paths with consideration of various real world constraints provides reasonable solution in a faster way than others.

• Transactions of the Korean Society of Mechanical Engineers A
• /
• v.33 no.12
• /
• pp.1401-1409
• /
• 2009

All the forces in the real world act dynamically on structures. Design and analysis should be performed based on the dynamic loads for the safety of structures. Dynamic (transient or vibrational) responses have many peaks in the time domain. Topology optimization, which gives an excellent conceptual design, mainly has been performed with static loads. In topology optimization, the number of design variables is quite large and considering the peaks is fairly costly. Topology optimization in the frequency domain has been performed to consider the dynamic effects; however, it is not sufficient to fully include the dynamic characteristics. In this research, linear dynamic response topology optimization is performed in the time domain. First, the necessity of topology optimization to directly consider the dynamic loads is verified by identifying the relationship between the natural frequency of a structure and the excitation frequency. When the natural frequency of a structure is low, the dynamic characteristics (inertia effect) should be considered. The equivalent static loads (ESLs) method is proposed for linear dynamic response topology optimization. ESLs are made to generate the same response field as that from dynamic loads at each time step of dynamic response analysis. The method was originally developed for size and shape optimizations. The original method is expanded to topology optimization under dynamic loads. At each time step of dynamic analysis, ESLs are calculated and ESLs are used as the external loads in static response topology optimization. The results of topology optimization are used to update the design variables (density of finite elements) and the updated design variables are used in dynamic analysis in a cyclic manner until the convergence criteria are satisfied. The updating rules and convergence criteria in the ESLs method are newly proposed for linear dynamic response topology optimization. The proposed updating rules are the artificial material method and the element elimination method. The artificial material method updates the material property for dynamic analysis at the next cycle using the results of topology optimization. The element elimination method is proposed to remove the element which has low density when static topology optimization is finished. These proposed methods are applied to some examples. The results are discussed in comparison with conventional linear static response topology optimization.

• IE interfaces
• /
• v.18 no.3
• /
• pp.297-307
• /
• 2005

PCB assembly is a complicated and difficult process to optimize due to the necessity of simultaneous consideration of component’s rack assignment and board mounting sequencing. An efficient PCB assembly method is developed by using neighboring positioned identical components information as well as quantity and size of the components. It is found that same type of components are located closely each other by checking real PCBs and interviewing with PCB designers in practice. Better performance of the developed procedure is obtained along with more number of total components and more number of neighboring positioned identical components cases. Simulation models are developed using Visual C++ for performance evaluation purposes of the suggested heuristic.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2006.11a
• /
• pp.285-288
• /
• 2006

Most previous studies on improving the effectiveness of CBR have focused on the similarity function aspect or optimization of case features and their weights. However, according to some of the prior research, finding the optimal k parameter for the k-nearest neighbor (k-NN) is also crucial for improving the performance of the CBR system. Nonetheless, there have been few attempts to optimize the number of neighbors, especially using artificial intelligence (AI) techniques. In this study, we introduce a genetic algorithm (GA) to optimize the number of neighbors that combine, as well as the weight of each feature. The new model is applied to the real-world case of a major telecommunication company in Korea in order to build the prediction model for the customer profitability level. Experimental results show that our GA-optimized CBR approach outperforms other AI techniques for this mulriclass classification problem.

• Communications for Statistical Applications and Methods
• /
• v.15 no.1
• /
• pp.125-136
• /
• 2008

In many practical engineering design problems, the form of objective functions is not given explicitly in terms of design variables. Given the value of design variables, under this circumstance, the value of objective functions is obtained by real/computational experiments such as structural analysis, fluid mechanic analysis, thermodynamic analysis, and so on. These experiments are, in general, considerably expensive. In order to make the number of these experiments as few as possible, optimization is performed in parallel with predicting the form of objective functions. Response Surface Methods (RSM) are well known along this approach. This paper suggests to apply Support Vector Machines (SVM) for predicting the objective functions. One of most important tasks in this approach is to allocate sample data moderately in order to make the number of experiments as small as possible. It will be shown that the information of support vector can be used effectively to this aim. The effectiveness of our suggested method will be shown through numerical example which is well known in design of engineering.

• Journal of Korean Society of Steel Construction
• /
• v.18 no.2
• /
• pp.213-223
• /
• 2006

Conventional truss optimization approaches, while often sophisticated and computationally intensive, have been applied to simple, minimum weight-cost models. These approaches do not perform well when applied to real-world trusses, which have costmodels that are complex and which often involve multiple objectives. Thus, this paper describes the optimization strategies that a clustering technique, which identifies members that are likely to have the same product type, uses for the optimal design of truss structures with real- world cost functions that consider the costs on the weight of the truss, the number of products in the design, the number of joints in the structures, and the costs required in the site.At first, the clustering technique is applied to identify the members and to generate a proper initial solution. A simple taboo search technique is then used, which attempts to generate the optimal solution by starting with the solution from the previous technique. For example, the proposed approach is a plied to a typical problem and to a problem similar to relative performances. The results show that this algorithm generates not only better-quality solutions but also more efficient ones

• Proceedings of the KAIS Fall Conference
• /
• 2003.11a
• /
• pp.178-184
• /
• 2003

Case-based reasoning (CBR) often shows significant promise for improving effectiveness of complex and unstructured decision making. Consequently, it has been applied to various problem-solving areas including manufacturing, finance and marketing. However, the design of appropriate case indexing and retrieval mechanisms to improve the performance of CBR is still challenging issue. Most of previous studies to improve the effectiveness for CBR have focused on the similarity function or optimization of case features and their weights. However, according to some of prior researches, finding the optimal k parameter for k-nearest neighbor (k-NN) is also crucial to improve the performance of CBR system. Nonetheless, there have been few attempts which have tried to optimize the number of neighbors, especially using artificial intelligence (AI) techniques. In this study, we introduce a genetic algorithm (GA) to optimize the number of neighbors to combine. This study applies the new model to the real-world case provided by an online shopping mall in Korea. Experimental results show that a GA-optimized k-NN approach outperforms other AI techniques for purchasing behavior forecasting.

• Journal of Korea Water Resources Association
• /
• v.43 no.1
• /
• pp.67-79
• /
• 2010

Korea water resources corporation (K-Water) has developed the real-time water resources management system for the Nakdong and the Geum River basin to efficiently operate multi-purpose dams in the basins. This study has extended to the Han River basin for providing an effective ending target storage of a month to the real-time water resources management system using Sampling Stochastic Dynamic Programming (SSDP), consequently increasing the efficiency of the reservoir system. The optimization model were developed for three reservoirs, named Soyang, Chungju, and Hwacheon, with high priority in terms of the amounts of effective capacity and water supply for the basin. The number of storage state variable for each dam to set an optimization problem has been assigned from the results of sensitivity analysis. Compared with the K-water operating policy with the target water supply elevations, the optimization model suggested in this study showed that the shortfalls are decreased by 37.22 MCM/year for the required water demands in the basin, even increasing 171 GWh in hydro electronic power generation. In addition, the result of a reservoir operating system during the drawdown period applied to real situation demonstrates that additional releases for water quality or hydro electronic power generation would be possible during the drawdown period between 2007 and 2008. On the basis of these simulation results, the applicability of the SSDP model and the reservoir operating system is proved. Therefore, the more efficient reservoir operation can be achieved if the reservoir operating system is extended further to other Korean basins.

• Journal of the Institute of Electronics Engineers of Korea TC
• /
• v.44 no.7 s.361
• /
• pp.112-121
• /
• 2007

In this paper, we propose a new checkpointing strategy for dual modular redundancy real-time systems. For every checkpoints the execution results from two processors, and the result saved in the previous checkpoint are compared to detect faults. We devised an operation algorithm in chectpoints to recover from transient faults as well as permanent faults. We also develop a Markov model for the optimization of the proposed checkpointing strategy. The probability of successful task execution within its deadline is derived from the Markov model. The optimal number of checkpoints is the checkpoints which makes the successful probability maximum.