Robust, interpretable and high quality fuzzy rule discovery using krill herd algorithm
In the paper we will try to generate fuzzy rule based systems (FRBSs) that follow three objectives: accuracy, interpretability and robustness. Accuracy is based on the Number of Patterns that are Correctly Classified (NCP), interpretability is measured as the Number of Rules (NR) and Sum of the Rules Length (SRL) and Robustness is measured as Sum of the accuracy, NR and SRL Standard Deviations (Sum of SDs) in successive runs. The algorithms that have high quality results with low SDs and in each run, the output is not being different, are called robust algorithms. Our algorithm consists of two stages based on the Krill Herd (KH) evolutionary algorithm; in the first stage the candidate rules are generated intelligently so in the second stage, those rules will be selected that get us closer to our objectives. Stage 2 of our algorithm can be used as a post processing algorithm on other algorithms and converts those to robust algorithms. The results show that, our algorithm has zero SDs with high accuracy and we have been successful in improving those three objectives that were in conflict.