FCANN: A new approach for extraction and representation of knowledge from ANN trained via Formal Concept Analysis

Abstract

Nowadays, Artificial Neural Networks (ANN) are being widely used in the representation of different systems and physics processes. Once trained, the networks are capable of dealing with operational conditions not seen during the training process, keeping tolerable errors in their responses. However, humans cannot assimilate the knowledge kept by those networks, since such implicit knowledge is difficult to be extracted. In this work, Formal Concept Analysis (FCA) is being used in order to extract and represent knowledge from previously trained ANN. The new FCANN approach permits to obtain a complete canonical base, non-redundant and with minimum implications, which qualitatively describes the process being studied. The approach proposed has a sequence of steps such as the generation of a synthetic dataset. The variation of data number per parameter and the discretization interval number are adjustment factors to obtain more representative rules without the necessity of retraining the network. The FCANN method is not a classifier itself as other methods for rule extraction; this approach can be used to describe and understand the relationship among the process parameters through implication rules. Comparisons of FCANN with C4.5 and TREPAN algorithms are made to show its features and efficacy. Applications of the FCANN method for real world problems are presented as case studies.