| Co-Investigator(Kenkyū-buntansha) |
KURODA Kenichi The University of Aizu, Professor, コンピュータ理工学部, 教授 (60285046)
HAMMAMI Ommar The University of Aizu, Assoc. Professor, コンピュータ理工学部, 助教授 (60254075)
ZHAO Qianfu The University of Aizu, Professor, コンピュータ理工学部, 教授 (90260421)
|
|---|
| Research Abstract |
The theoretic foundation of this research is a Society model for cooperative co-evolutionary algorithms (CoopCEAs), which was proposed currently by Zhao. CoopCEA is a new paradigm for evolutionary computation, which is, as believed by many researchers, more suitable for solving large-scale and complex problems. Our final goal is to realize the society model in hardware, so that practical systems can be evolved in lightning speed. This goal, however, is too difficult to be achieved in two years. One reason is that there are still a few open problems to be solved concerning the societv model. The second reason is that current hardware (say, FPGA) technology is still not sufficient for realizing such a genetic breeder.
Based on these considerations, we have focused our attention on two topics. One is to make the society model more theoretically sound and more hardware implementable. Another is to implement a simplified society model, and apply it to solve a simple pattern recognition probl
… More
em. These two topics were studied in parallel.
As for the first part, we modified our source-code from C to C++, so that all objects are realized by software components, which can be easily modified and implemented by hard components in the future. We then studied the evaluation of the cooperative modules, which is the core problem to be solved in using the society model. Although we got some results through experiments, we are still not clear how to use CoopCEA more effectively in general. Another key point is how to decompose a large system into many modules. In general, we do not have to decompose the system explicitly in using the society model, because in general, the CoopCEAs can breed systems using yet-to-be-evolved components and yet-to-be-evolved structures. In practice, however, if we can decompose the system properly, the evolutioncan be greatly accelerated. For this purpose, we studied evolutionary learning of decision trees and nearest
neighbor multilayer perceptrons. We have just got some important clues. Interesting results are expected to be reported in near future.
As for second part, we implemented a simple nearest neithbor based neural network usint FPGA (Altera Flex10K100). It is connected through PCT bus to a host computer. All genetic operations are performed in the computer, and the evaluation part is performed on the FPGA chip. The chip, with minor revision,
could be used for the simple Co-Evolutionary algorithm. This will be verified in the next step. Less
|
