• Journal of Korea Multimedia Society
• /
• v.21 no.8
• /
• pp.969-981
• /
• 2018

Recent advances in machine learning approaches such as deep neural network and reinforcement learning offer significant performance improvements in generating detailed and varied motions in physically simulated virtual environments. The optimization methods are highly attractive because it allows for less understanding of underlying physics or mechanisms even for high-dimensional subtle control problems. In this paper, we propose an efficient learning method for stochastic policy represented as deep neural networks so that agent can generate various energetic motions adaptively to the changes of tasks and states without losing interactivity and robustness. This strategy could be realized by our novel trajectory search method motivated by the trust region policy optimization method. Our value-based trajectory smoothing technique finds stably learnable trajectories without consulting neural network responses directly. This policy is set as a trust region of the artificial neural network, so that it can learn the desired motion quickly.

• Journal of the Institute of Convergence Signal Processing
• /
• v.13 no.4
• /
• pp.207-210
• /
• 2012

Genetic algorithms(GA) are well known and very popular stochastic optimization algorithm. Although, GA is very powerful method to find the global optimum, it has some drawbacks, for example, premature convergence to local optima, slow convergence speed to global optimum. To enhance the performance of GA, this paper proposes an adaptive partitioning-based genetic algorithm. The partitioning method, which enables GA to find a solution very effectively, adaptively divides the search space into promising sub-spaces to reduce the complexity of optimization. This partitioning method is more effective as the complexity of the search space is increasing. The validity of the proposed method is confirmed by applying it to several bench mark test function examples and the optimization of fuzzy controller for the control of an inverted pendulum.

• Proceedings of the IEEK Conference
• /
• 2001.06c
• /
• pp.13-16
• /
• 2001

In reinforcement teaming, Q-learning converges quite slowly to a good policy. Its because searching for the goal state takes very long time in a large stochastic domain. So I propose the speedup method using the Q-value initialization for model-free reinforcement learning. In the speedup method, it learns a naive model of a domain and makes boundaries around the goal state. By using these boundaries, it assigns the initial Q-values to the state-action pairs and does Q-learning with the initial Q-values. The initial Q-values guide the agent to the goal state in the early states of learning, so that Q-teaming updates Q-values efficiently. Therefore it saves exploration time to search for the goal state and has better performance than Q-learning. 1 present Speedup Q-learning algorithm to implement the speedup method. This algorithm is evaluated. in a grid-world domain and compared to Q-teaming.

• Journal of the Korean Society of Safety
• /
• v.7 no.2
• /
• pp.5-13
• /
• 1992

If zero-accident movement is to be successful, the objective goal period should be surely obtainable, and much more in our country where frequency rate of injury are remarkably fluc-tuating. However In our country, as far as we know, no method to establish a reasonable zero-accident goal period is guaranteed. In thls paper, a new establishing-method of reasonable goal period for individual industry with considering recent accident trend is presented. A mathematical model for industrial accidents generation was analyzed, and a stochastic process model for the accident generation inteual was formulated. This model could tell the accident generation rate in future by understanding the accident tendency through the time-series analysis and search for the distribution of numbers of accidents and accident interval. On the basis of this, the forecasting method of goal achievement probability by the size and the establishment method of reasonable goal period were developed.

• Structural Engineering and Mechanics
• /
• v.81 no.4
• /
• pp.429-441
• /
• 2022

A probabilistic assessment of the seismic-excited buildings with a multiple-tuned-mass-damper (MTMD) system is carried out in the presence of uncertainties of the structural model, MTMD system, and the stochastic model of the seismic excitations. A free search optimization procedure of the individual mass, stiffness and, damping parameters of the MTMD system based on the snap-drift cuckoo search (SDCS) optimization algorithm is proposed for the optimal design of the MTMD system. Considering a 10-story structure in three cases equipped with single tuned mass damper (STMS), 5-TMD and 10-TMD, sensitivity analyses are carried out using Sobol' indices based on the Monte Carlo simulation (MCS) method. Considering different seismic performance levels, the reliability analyses are done using MCS and kriging-based MCS methods. The results show the maximum structural responses are more affected by changes in the PGA and the stiffness coefficients of the structural floors and TMDs. The results indicate the kriging-based MCS method can estimate the accurate amount of failure probability by spending less time than the MCS. The results also show the MTMD gives a significant reduction in the structural failure probability. The effect of the MTMD on the reduction of the failure probability is remarkable in the performance levels of life safety and collapse prevention. The maximum drift of floors may be reduced for the nominal structural system by increasing the TMDs, however, the complexity of the MTMD model and increasing its corresponding uncertainty sources can be caused a slight increase in the failure probability of the structure.

• Structural Engineering and Mechanics
• /
• v.24 no.5
• /
• pp.529-541
• /
• 2006

This work gives an application of stochastic techniques for the optimization of stiffened plates in vibration. The search strategy consists of substituting, for finite element calculations in the optimization process, an approximate response from a Rayleigh-Ritz method. More precisely, the paper describes the use of a Rayleigh-Ritz method in creating function approximations for use in computationally intensive design optimization based on genetic algorithms. Two applications are presented; their deal with the optimization of stiffeners on plates by varying their positions, in order to maximize some natural frequencies, while having well defined dimensions. In other words, this work gives the fundamental idea of using a Ritz approximation to the response of a plate in vibration instead of finite element analysis.

• Proceedings of the KIEE Conference
• /
• 2002.07a
• /
• pp.121-123
• /
• 2002

In this paper, we preposed a chaos optimization method to reduce computational effort and enhance optimality of the solution in feeder reconfiguration problem. Chaos method in optimization problem searches the global optimal solution on the regularity of chaotic motions and more easily escapes from local or near optimal solution than stochastic optimization algorithms. The chaos optimization method is tested on 15 buses and 32 buses distribution systems, and the test results indicate that it is able to determine appropriate switching options for global optimum configuration with less computation.

• Journal of the Korean Society of Marine Environment & Safety
• /
• v.25 no.3
• /
• pp.259-268
• /
• 2019

Even if only two ships are encountered, a collision may occur due to the mistaken judgment of the positional relationship. In other words, if an officer does not know a target ship's intention, there is always a risk of collision. In this paper, the experiments are conducted to investigate how the intention affects the action of collision avoidance in cooperative and non-cooperative situations. In non-cooperative situation, each ship chooses a course that minimizes costs based on the current situation. That is, it always performs a selfish selection. In a cooperative situation, the information is exchanged with a target ship and a course is selected based on this information. Each ship uses the Distributed Stochastic Search Algorithm so that a next-intended course can be selected by a certain probability and determines the course. In the experimental method, four virtual ships are set up to analyze the action of collision avoidance. Then, using the actual AIS data of eight ships in the strait of Dover, I compared and analyzed the action of collision avoidance in cooperative and non-cooperative situations. As a result of the experiment, the ships showed smooth trajectories in the cooperative situation, but the ship in the non-cooperative situation made frequent big changes to avoid a collision. In the case of the experiment using four ships, there was no collision in the cooperative situation regardless of the size of the safety domain, but a collision occurred between the ships when the size of the safety domain increased in cases of non-cooperation. In the case of experiments using eight ships, it was found that there are optimal parameters for collision avoidance. Also, it was possible to grasp the variation of the sailing distance and the costs according to the combination of the parameters, and it was confirmed that the setting of the parameters can have a great influence on collision avoidance among ships.

• Computational Structural Engineering
• /
• v.10 no.4
• /
• pp.349-358
• /
• 1997

The present paper describes an investigation into the application of the genetic algorithm (GA) in the optimum design of composite laminated structure. Stochastic processes generate an initial population of designs and then apply principles of natural selection/survival of the fittest to improve the designs. The five test functions are used to verify the robustness and reliability of the GA, and as a numerical example, minimum weight of a cantilever composite laminated beam with a mix of continuous, integer and discrete design variables is obtained by using the GA with exterior penalty function method. The design problem has constraints on strength, displacements, and natural frequencies, and is formulated to a multidimensional nonlinear form. From the results, it is found that the GA search technique is very effective to find the good optimum solution as well as has higher robustness.

• Journal of the Korea Society of Computer and Information
• /
• v.6 no.1
• /
• pp.82-89
• /
• 2001

It is a document segmentation to set a boundary in the documents by the contents. It is essential for the accurate and efficient information search. In this paper we want to make an automatic document segmentation system with the method of probability analysis which uses the mutual information between the words. Proposed system can move the boundary of window and compute the similarity or the two window. In this system the more words are shared and the more important the words are, the higher the cohesive force of the two window systems goes. The result of experience with the document segmentation is that despite the differences of block unit the division point at which we expected to divide was normally divided.