Some Possibilities of Improving the CORA Classification Algorithm

Abstract

This paper examines some possibilities of improving the CORA classification algorithm developed by M. Bongard in sixties. The algorithm is based on finding features of objects one needs to classify. The theoretical part of this study explains two main shortcomings of the CORA classification algorithm: (1) the algorithm rejects features that at least once appear in the opposite class and thus loses potentially valuable information; and (2) the algorithm has extremely large learning time that can be due to the “combinatorial explosion”. The study suggests two methods to overcome these difficulties and to improve the algorithm: (1) the method of relatively good features and (2) the method of sequential covering. The experimental results demonstrate that both the methods suggested ensure a sufficient algorithm performance’s improvement as compared to the classic Bongard algorithm implementation. Best results are achieved when both methods are combined and sufficiently improve the classification precision and reduce learning time.