| Budget Amount *help |
¥3,210,000 (Direct Cost: ¥3,000,000、Indirect Cost: ¥210,000)
Fiscal Year 2007: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
Fiscal Year 2006: ¥600,000 (Direct Cost: ¥600,000)
Fiscal Year 2005: ¥1,200,000 (Direct Cost: ¥1,200,000)
Fiscal Year 2004: ¥500,000 (Direct Cost: ¥500,000)
|
|---|
| Research Abstract |
A purpose of this research is to obtain a model with high generalization ability in supervised learning by structural learning algorithms, which optimize the model parameters and the model structure simultaneously based on some ideas taken from degeneration phenomena. In the structure learning algorithms based on degeneration, a gene is represented by a pair of a normal value and its damage rate. In the algorithms, damaging mutation to change a gene toward a more damaged gene by increasing the damaged rate is introduced in addition to crossover for genes and mutation for normal values. The structural learning is realized by assuming completely damaged genes as useless model parameters and deleting the parameters. We successfully proposed the following structural learning algorithms based on degeneration: (1) GAd, which realizes structure learning by one-point crossover of genes, mutation for normal values, and damaging mutation for damage rates, based on genetic algorithm, (2) CGAd, in
… More
which two populations, the population of control individuals to control damaging mutation and the population of learning individuals to realize structural learning based on degeneration, co-evolve to adjust damaging mutation adaptively, (3) Ded, which realizes genetic operations of Differential Evolution by using an mapping from the genotype (represented by a normal value and its damaged rate) to the phenotype, and inverse mapping from the phenotype to the genotype. It has been shown that the adjustment of the degeneration pressure is very important. When the pressure is too large, the learning ability becomes insufficient. When the pressure is too small, the model structure is not optimized enough. Because adjusting the degeneration pressure is difficult for Gad, we proposed CGAd which can control the pressure adaptively by coevolution and Ded which can obtain a model structure with high learning ability even in high degeneration pressure. In future, we plan to control the degeneration pressure adaptively by introducing coevolution into Ded and continue to study the prediction of time series by using the structural leaning algorithms based on degeneration. Less
|
