Soft Computing in Artificial Intelligence: Uses, Directions, and Future Prospects
Daniel J. Buehrer · 1994
This review briefly sketches the history of ”soft computing”, which encompasses the fields of fuzzy logic, neural networks, and probabilistic reasoning. The field is characterized by the use of numbers (usually between 0 and 1) to represent weights or fuzzy classifications. These weights can, in parallel, find the most suitable classification of an input consisting of various feature values, thus leading to the most appropriate output. The model for combining these values can either be a simple fuzzy rule-based model, a self-adaptive neural network, or a genetically evolving rule-base. The inherent parallelism and the self-learning nature of soft computing holds the potential to break the limitations which have traditionally held back binary-logic VonNeumann machines. Perhaps even more important than the potential parallelism, however, is the paradigm shift in the nature of science which has been brought about by soft computing. Previously, scientific theories had to be precise. Now it is recognized by some scientists that it is impossible to be precise about complex systems like economics or sociology, or even fairly simple non-linear systems involving several variables. For such systems, fuzzy models cannot be disproven like traditional scientific theories. It can only be shown that one model seems to work better than another model for a given test suite.