Framework development in extracting rules from trained neural network / Shuzlina Abdul Rahman, Azlinah Hj Mohamed and Marina Yusoff

Rules Extraction (RE) technique has become a significant role in the area of Artificial Neural Network (ANN). It can facilitate ANN developer and most crucially it helps its user to generate symbolic rules that enable them to understand the knowledge inside it in an explainable form. There are many RE techniques have been explored and tested by several researchers in different domains. This report presents a general framework for RE techniques classification, which focuses on three approaches namely decompositional, pedagogical and eclectic. In addressing this framework, the criteria of each approaches has been explored and analyzed from eight factors: process extraction, merit, demerit, rule type, type of data, rule quality, processing complexity, and the description of each RE technique. The analysis is derived by excavating literature on RE techniques starting from the year 1993 until 2003 focused on supervised learning algorithm. The framework primarily demonstrates that each approach does not require a special training process for a new input dataset and does not require special network architecture and it can be used as a guideline to ANNs researcher and developer to choose a suitable RE techniques in order for them to perform ANNs' research or developing ANN applications.