Fuzzy Rule based systems : applications, issues, solutions and open problems - where do we stand?

Nikhil R. Pal
Indian Statistical Institute

Fuzzy rule based systems have been used in numerous applications from washing machines to missile guidance. Two very distinct advantages of fuzzy rule based systems over other tools like neural networks are : they are readable and are less likely to produce poor generalization due to the reasoning paradigm built into it. Yet, when it comes to solving application problems (pattern recognition type) by a practitioner, more often neural networks become the tool of choice. Why? An attempt to find the reasons behind this revealed the following answers/points : “fuzzy systems are difficult to design”, “no good tool box, like Matlab neural networks toolbox, is available”, “may not work for complex problems” …. So, where do we really stand and how relevant these points are?

We shall motivate our discussion with four representative applications where fuzzy rule based systems have done a great job. There four problems are : structure preserving dimensionality reduction, land-cover classification, detection of bounded weak echo region, and cloud-cover estimation. Then we shall identify and discuss various questions/issues that one needs to address for designing a useful fuzzy rule based system. Our discussion will be restricted to automatic methods based on exploratory data analysis. The list of issues will include: what should we cluster, how to decide on the optimal number of rules, is use of cluster validity appropriate, should we use the same reasoning scheme in all contexts, how to achieve readability, and many other interesting issues. We shall discuss some solutions to each issue and indicate how other choices can provide rich families of solutions. In spite of tremendous advancements in fuzzy rule based systems identification, there are a few problems that often make rule finding difficult and also limit the applicability of fuzzy rule based systems – we shall finish with a discussion on these open problems.