A New Adaptive Random Search Method in Neural Networks
-
- HIRASAWA Kotaro
- Graduate School of Information Science and Electrical Engineering, Kyushu University
-
- TOGO Kazuyuki
- Graduate School of Information Science and Electrical Engineering, Kyushu University
-
- HU Jinglu
- Graduate School of Information Science and Electrical Engineering, Kyushu University
-
- OHBAYASHI Masanao
- Graduate School of Information Science and Electrical Engineering, Kyushu University
-
- SHAO Ning
- Graduate School of Information Science and Electrical Engineering, Kyushu University
-
- MURATA Junich
- Graduate School of Information Science and Electrical Engineering, Kyushu University
Bibliographic Information
- Other Title
-
- ニューラルネットワークの適応的ランダム探索最適化手法
- RasID
- RasID
Abstract
In this paper, a new random optimization method is presented. The proposed method is called RasID which is an abbreviation of Random Search with Intensification and Diversification and it can search for a global minimum based on the probability density function of the searching, which can be modified using information based on success or failure of the past searching to execute intensified and diversified searching. A principle of the RasID is such that searching for the parameters is strengthened when the probability of finding good solutions is high, and is weakened when the probability is low. In other words, when the searching is success, the probability density function is modified so that the search is done within a short range to the direction of decreasing the criterion function in order to achieve a certain improvement of the criterion function. On the other hand, when the search is failure, the searching range is expanded and the direction of the searching is adjusted to include even those which make the criterion function worse in order to search the parameters with a wide range. The RasID is superior to the commonly used random search method because the RasID is a kind of adaptive random search utilizing the past experiences of the searching as mentioned above. From the simulation results of realizing many kinds of nonlinear functions and controlling a nonlinear crane system using Neural Networks (N.N.), it is clarified that the RasID is superior in performances to the Gradient Method.
Journal
-
- Transactions of the Society of Instrument and Control Engineers
-
Transactions of the Society of Instrument and Control Engineers 34 (8), 1088-1096, 1998
The Society of Instrument and Control Engineers
- Tweet
Details 詳細情報について
-
- CRID
- 1390001204501611648
-
- NII Article ID
- 130003791437
-
- ISSN
- 18838189
- 04534654
-
- Data Source
-
- JaLC
- Crossref
- CiNii Articles
-
- Abstract License Flag
- Disallowed