Weighted Fuzzy Production Rule Extraction Using Modified Harmony Search Algorithm and BP Neural Network Framework

Hang-Cheng Li,Feng Qin,Azlan Mohd Zain,Li-Ping Mo,Kai-Qing Zhou

doi:10.1109/access.2020.3029966

Abstract

Compared with rules in the form of `IF-THEN,' weighted fuzzy production rules (WFPRs) have more robust knowledge expression capabilities, but weighted fuzzy production rules are more difficult to obtain. The weighted fuzzy production rules obtained using traditional neural network methods have shortcomings, such as insufficient precision and insufficient knowledge extraction. Focusing on the mentioned shortages, a modified weighted fuzzy production rules extraction approach is proposed by combining the modified harmony search algorithm, and neural network. The method consists of three main stages. First, a global optimal adaptive harmony search algorithm (AGOHS) is proposed to overcome the traditional harmony search algorithm's existing poor adaptive ability. Then, the AGOHS algorithm is used to optimize the neural network's initial weights to improve the neural network's training efficiency. Finally, extract the WFPRs with IF-THEN from the trained neural network and give the corresponding fuzzy reasoning. Through the WFPRs extraction experiments using IRIS and PIMA data sets reveal the proposed rule extraction framework has some apparent highlights, such as high accuracy, the smaller number of generated rules, and low redundancy.

Highlights

Production Rule with ‘IF-’ formalism is one of the most common representation forms of knowledge in the field of artificial intelligence, which has the advantages to understand and to add and to delete or to update related information [1]
By inspired by the existing rule extraction approach using NN and Soft computing techniques, this paper proposes a method of extracting weighted fuzzy production rules using hybrid neural networks and improved harmony search algorithm to solve the classification issue of continuous realvalued attributes
The fuzzy reasoning process can be understood as the process of converting input sample attribute values into corresponding categories using weighted fuzzy production rules. This process is similar to the process of neural network calculation of output values, while the weighted fuzzy production rules are generated based on the weights of the neural network connection, so the reasoning method is generated by simulating the flow of data from the input node to the output node of the neural network

Summary

INTRODUCTION

Production Rule with ‘IF-’ formalism is one of the most common representation forms of knowledge in the field of artificial intelligence, which has the advantages to understand and to add and to delete or to update related information [1]. Li et al.: WFPR Extraction Using Modified Harmony Search Algorithm and BP Neural Network Framework extracts rules from trained neural networks [9]. The AGOHS algorithm proposed in this paper has good global optimization and self-adaptive capabilities In theory, it can perform a good error optimization on the objective function of the neural network to improve the knowledge accuracy of the extracted rules. By inspired by the existing rule extraction approach using NN and Soft computing techniques, this paper proposes a method of extracting weighted fuzzy production rules using hybrid neural networks and improved harmony search algorithm to solve the classification issue of continuous realvalued attributes. Li et al.: WFPR Extraction Using Modified Harmony Search Algorithm and BP Neural Network Framework TABLE 1 Unknown samples can be simulated and predicted

STRUCTURAL LEARNING WITH FORGETTING OF NN ALGORITHM

WEIGHTED FUZZY PRODUCTION EXTRACTION BY AGOHS AND BP NEURAL NETWORK

CORRESPONDING FUZZY REASONING METHOD

EXPERIMENTS AND RELATED ANALYSES

IRIS DATABASE EXPERIMENT

CONCLUSION