Abstract
Fuzzy rules and inferences provide a powerful framework for controlling complex processes.
Like symbolic AI systems, they offer a comprehensible representation, which facilitates transfer
of knowledge between the system and human experts and users. Moreover, unlike most symbolic
systems, the responses are more gradual. In this fuzzy representation, knowledge is distributed
among three different levels: symbolic rules, numerical weights, and fuzzy linguistic
definitions. When human expertise is used to build a fuzzy controller, it must often be tuned to
fit particular objectives or simply be corrected when it is incomplete or otherwise incorrect.
When such expertise does not exist or is difficult to acquire and model in the controller, the
knowledge must be generated. In either case, it is desired to perform these tasks automatically.
This can be accomplished by searching the space of all controllers, while sampling their performance
to provide heuristic quality measures. This requires a robust search mechanism. Moreover,
the search should be conducted simultaneously at all representation levels to avoid any
assumptions about the knowledge. Genetic algorithms provide the necessary robustness. Their
previous applications concentrated on manipulations of a single knowledge level. In this paper,
we investigate their applications to manipulations of two different levels: the symbolic rules and
the numerical weights. This is accomplished adapting a standard numerical genetic algorithm.
A number of experiments is conducted and reported, which illustrate both tuning and learning
capabilities of the system.
| Original language | American English |
|---|---|
| Journal | Proceedings of the 1994 ACM Symposium on Applied Computing |
| State | Published - 1994 |
Disciplines
- Computer Sciences