Rule base compression in fuzzy systems by filtration of non-monotonic rules

This paper proposes a rule base compression method for Mamdani fuzzy systems with non-monotonic rules. The method is based on filtration of non-monotonic rules whereby the redundant computations in the fuzzy inference with respect to the crisp values of the inputs to the fuzzy system are removed. The method identifies all redundant rules after fuzzification and removes them while preserving the defuzzified output from the fuzzy system for each simulation cycle. In comparison to other rule base reduction methods, this method does not compromise the solution and is more efficient in terms of on-line computations within a wide operating range. The method processes the rule base during simulation cycles by contracting it to a rule base of a smaller size at the start of each inference stage and then expanding it to its original size before the next fuzzification stage.