A case for learning simpler rule sets with multiobjective evolutionary algorithms

Abstract

Fuzzy rules can be understood by people because of their specification in structured natural language. In a wide range of decision support applications in business, the interpretability of rule based systems is a distinguishing feature, and advantage over, possible alternate approaches that are perceived as "black boxes", for example in facilitating accountability. The motivation of this paper is to consider the relationships between rule simplicity (the key component of interpretability) and out-of-sample performance. Forecasting has been described as both art and science to emphasize intuition and experience aspects of the process: aspects of intelligence manifestly difficult to reproduce artificially. We explore, computationally, the widely appreciated forecasting "rule-of-thumb" expressed in Ockham's principle that "simpler explanations are more likely to be correct".