다구찌 실험계획법을 활용한 LSTM의 하이퍼파라미터 최적화에 관한 연구 (A Study on the Optimization of Long Short-Term Memory Hyperparameters Using the Taguchi Design of Experiments)

서지정보

· 발행기관 : 한국신뢰성학회
· 수록지 정보 : 신뢰성 응용연구 / 21권 / 3호 / 238 ~ 245페이지
· 저자명 : 이승훈, 김용수

초록

Purpose: Herein, we propose a hyperparameter optimization methodology for long short-term memory (LSTM) using the Taguchi design of experiments.
Methods: After preprocessing the data, a manual search was used to select the factor levels. The grid search method was applied to four levels and five factors, yielding results of 1024 combinations. Based on the results, the Taguchi method was used to obtain a total of 16 simple orthogonal arrays of four levels. The root mean squared error (RMSE) was obtained through three iterations, and the corresponding mean and standard deviation were evaluated. Furthermore, significant results were obtained when the proposed methodology was compared to the grid search methodology.
Results: We confirmed the significance of the proposed methodology through a comparative analysis of the results of the grid search methodology, which performed combinations for all factors and levels. When the proposed methodology was used, the accuracy of the grid search methodology improved, and the time required to perform parameter optimization was significantly reduced.
Conclusion: According to this study, a level of results significantly reduces the number and time of measurements when the Taguchi method is used, and it does not differ from the resulting values when both are executed.

영어초록

Purpose: Herein, we propose a hyperparameter optimization methodology for long short-term memory (LSTM) using the Taguchi design of experiments.
Methods: After preprocessing the data, a manual search was used to select the factor levels. The grid search method was applied to four levels and five factors, yielding results of 1024 combinations. Based on the results, the Taguchi method was used to obtain a total of 16 simple orthogonal arrays of four levels. The root mean squared error (RMSE) was obtained through three iterations, and the corresponding mean and standard deviation were evaluated. Furthermore, significant results were obtained when the proposed methodology was compared to the grid search methodology.
Results: We confirmed the significance of the proposed methodology through a comparative analysis of the results of the grid search methodology, which performed combinations for all factors and levels. When the proposed methodology was used, the accuracy of the grid search methodology improved, and the time required to perform parameter optimization was significantly reduced.
Conclusion: According to this study, a level of results significantly reduces the number and time of measurements when the Taguchi method is used, and it does not differ from the resulting values when both are executed.

참고자료

· 없음