1993 Fiscal Year Final Research Report Summary
The Analysis of Learning Methods, Learning Speed, Learning Capacity for the Recurrent Type Neural Networks and their Applications
| Project/Area Number |
03452172
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
情報工学
|
|---|
| Research Institution | KANSAI UNIVERSITY (1992-1993)
Osaka City University (1991) |
|---|
Principal Investigator |
YAMASHITA Kazumi KANSAI UNIVERSITY, Engineering, Professor, 工学部, 教授 (40046850)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
OKAMOTO Jiro OSAKA CITY UNIVERSITY, Engineering, Assistant Professor, 工学部, 助教授 (30047146)
|
|---|
| Project Period (FY) |
1991 – 1993
|
| Keywords | neural network / recurrent type nerual network / parity discrimination / genetic algorithms / GA / learning / data flow / textile feeling |
|---|
| Research Abstract |
In the following, we summarize the research results.
(1) Using Genetic Algorithms as a learning method for the recurrent type NN's, there is no need to consider the problems of learning speed in the case of the learning being progressed, though usually the learning speed decreases, nor need to care about stability of the NN's, though they usually become unstable.
(2) We proposed a method to determine the hidden unit numbers automatically using GA into forward type NN's and demonstrated that the method is very useful to determine its numbers by experiments.
(3) We proposed a simple learning method for the recurrent type NN's networks using GA and proved that the method is very effective to synthesize spatial filters.
(4) We proved that the method using GA is effective for recognizing patterns of numbers by forward type NN's, compared with the method using back propagaton (BP) algorithm.
(5) We proved that the scheduling problems of data flow computers can be easily solved by the energy minimizing principle of Hopfield type NN's, and the method we proposed is more effective than the conventional one.
(6) We proved that the minimum number of hidden units which is necessary to discriminate the parity of N input bits is at most (N+1)/2.
(7) We demonstrated that the expert system to obtain the textile feeling "FUAI" can be realized by using the forward type NN's. Further based on the observation, we proposed and demonstrated that it is possibe to obtain physical factors from FUAI factors reversely.
(8) We proposed an algorithm to discriminate the amplitudes and phases of signals with very similar spectra in the case of low SN ratio and demonstrated that it is more accurate than the conventional ones.
|
