I Introduction
Computational intelligence (CI), as a rapid growing technology has nowadays substituted traditional artificial intelligence (AI) and expert systems in a wide range of applications. The main advantage of these new techniques over AI’ s symbolic processing, lies in the ability of the CI systems to adapt to the problem environment, hence bypassing the stage of human knowledge acquiring - a mandatory step using expert systems. How ever, in a number of high-level decision tasks, common expert systems remain still applicable. The reason can be noticed into the need for symbolic representation of the knowledge into these systems, which is a feature that many CI systems have unremarkable success. In other words, it is considered that symbolic representation can be of significant value in these systems for humans, by making clear the inference process to users. Among CI methodologies, neural networks are powerful connectionist systems that still lack the element of complete and accurate interpretation into human-understandable form of knowledge and remain a black box for experts. To heal this situation, a number of alternative approaches have been proposed. Neural logic networks [1] belong to this category, and by their definition can be interpreted into a number of Prolog rules that consist an expert system. Virtually every logic rule can be represented into these networks and then transformed into Prolog commands. Although this model offers excellent results when used within the AT framework (i.e. building a system in a top-down process), the application of neural logic networks in CI's data mining tasks considered a bottom-up procedure-has undergone limited success. The reason lies in that proposed systems suffered at least one of the following limitations:
The extracted neural logic network cannot be interpreted into expert rules [1]–[2].
The proposed methodology cannot express neural logic networks in their generic graph form [3].
The user has to select topology and network connection model [1]–[2].