![]() |
|||
![]()
|
![]() |
![]() |
![]() |
4. RESEARCH ISSUESUsing GAs and expert systems by themselves are well-established AI problem-solving techniques. Using them together, however, is not. GAs are computationally oriented, weak methods that do not need to know anything about the problem. Expert systems, on the other hand, are symbolically based with a structure that primarily depends on domain knowledge. Since expert system understanding is assumed of the reader, this section will primarily discuss GA research issues. The section explained that GAs use an evolutionary approach for generating new solutions. All of us understand the basic Darwinian concept; however, there are variations in the biological world that can be simulated using GAs for possible improvements. Bacterial genetics provides one such world. Bacteria can transfer DNA to recipient cells through mating. The female recipient cells can then change to male cells, thus facilitating the transfer of the acquired DNA to the entire bacteria population. This approach has been implemented to find fuzzy rules for a semi-active suspension controller (Hashiyama, Furuhashi, and Uchikawa, 1995). There are possibly other variations in the biological world that can be mimicked in order to improve the effectiveness of GAs. The knowledge representation issue is an important consideration in any type of AI implementation. In GAs, the traditional binary encoding can be replaced by other encodings depending on the problem domain. Gray codes were mentioned as one alternative. There presently exists numerous variations that were developed for a particular problem. Take, for example, the traditional traveling salesman problem. This problem can be solved using GAs. However, since each genotype in this case would represent a route between cities, it would be inefficient to have genotypes consisting of a sequence of numbered cities, since the crossover operation could create many invalid routes (duplication of cities, etc.). There are numerous ways to solve this representation problem. One generic implementation is to use random-key encoding. Each genotype would consist of a sequence of random numbers. To decode the genotype, the positioning of the genes would be decoded as the city number. For example, the genotype 6-1-4 would decode into the route city 2 --> city 3 --> city 1. It is up to the user's creativity to implement an encoding that would work best for their problem domain. GAs are concerned with adaptive behavior of populations of genotypes. Taking that concept to the next level results in a GA whose overall structure can adapt to the problem. Not only are the genotypes evolving, but also the structural components of the GA: fitness function, selection operators, and the search operators. As the GA works through many generations, information could be gathered about the problem which can be used to focus the GA toward better solutions. This can be called co-evolution or multilevel evolution, and research has been done on such systems (Wang, Goodman, and Punch, 1995). This type of GA model holds many promises and requires more research to fully explore its potential. 5. FUTURE TRENDS AND SUMMARYOne of the major trends is the subject of this article: hybrid systems. Researchers are constantly exploring different methods of combining AI tools to solve problems. These efforts attempt to combine different systems to take advantage of the strengths of different tools. Ideally, the strengths of one of the combined tools makes up for the weaknesses of the other. Exploring very large search spaces with GAs can be computationally taxing. The continual developments in computers and processing power allows researchers to tackle larger problems. Work is being done in the field of distributed or parallel GAs, which would take advantage of multiple platforms. Other research trends include explorations into multilevel evolution and finding the optimum parameter values for a given problem. Another theme that has been arising with increasing frequency is the inclusion of self-adapting mechanisms with GAs to control parameters involving the internal representation, mutation, recombination, and population size (DeJong, 1994). There has been much work since John Holland's initial research of genetic algorithms. Twenty years of additional research have led to various ideas and implementations in an effort to take advantage of the concept of evolution in the computer science field. The next 20 years should be just as interesting, as greater numbers of researchers work in this field and the tools available to them increase in speed and efficiency. REFERENCES
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |