Research On Methods Of Fuzzy Rules' Extraction Based On Concept Learning

Concept learning which learns knowledge from signed training data is the essential issue of inductive learning. With the wider implication of rules extraction in expert system, pattern recognition, image processing and sound recognition, rules extraction algorithms based on concept learning have interested more and more people.This paper firstly overviews the two classical inductive learning algorithms: AQ algorithm and decision tree algorithm. Some insufficiencies of them in both theory and practice are indicated. Then the work we have done in the field of rules extraction by inductive learning methods is introduced particularly. 1. A fuzzy implication operator (called NFIO) based on sub threshold, vectors' distance et al and a fuzzy rules extraction algorithm based on NFIO (called RENFIO) are put forward. The learning process of the rules is made into two steps by RENFIO: (1) Get the combination of each class's description which has the larger implication. (2) Find the best description of each class, and that is finding the best rales cover the training data. Experiments show that NFIO gives a better description of implication relationship between concepts. Compared to others, the advantages of RENFIO algorithm are: search concepts quickly; precision of prediction for new samples is higher; the rales are easier. 2. This paper also gives fuzzy Candidate-elimination algorithm to extract rales using both positive and negative instances. In the concept space constructed by features, we find a good description that can cover positive instances and exclude negative instances. How to learn in the feature hypothesis space is an important issue of concept learning. AQ algorithm and decision algorithm are two famous methods that can deal with the problem. Compared to them, fuzzy Candidate-elimination algorithm has the following advantages: search concept space for only once and so this algorithm can cope with large data set; has a stronger ability of classification as to continuous data; it can deal with noisy data. RENFIO algorithm's efficiency is testified by iris dataset and wine dataset.3. We also give three applications of NFIO: features extraction; obtaining each instance's description; gaining wrong classified instances' description. Experiments on three datasets have proved that the features extracted by NFIO are the most useful ones for classification and a best feature also can be found. Each instance's description found by NFIO is nearly the same as the one in chapter three and that illuminates the lightness of this method. Wrong classified instances are the ones not covered by fuzzy rules. Their description can complement the fuzzy rules and improve the accuracy of training data.