Development of an Optimal Entropy Classifier and Prudent Learning Model
Jyotsana Grover, Madasu Hanmandlu · IEEE Transactions on Artificial Intelligence · 2021
This article gives the representation of both probabilistic uncertainty and possibilistic certainty including the Bayesian learning in the framework of information set theory which is an offshoot of the Hanman–Anirban entropy function. Being information theoretic and parametric, this function deals with both probability and possibility. If a set of information source (attribute) values is fitted with this entropy function it gives rise to information values and the sum of these values is certainty. An adaptive form of this function yields the Hanman transform (HT) that gives the higher order certainty. An optimal entropy classifier is developed by learning the weight (support) vectors of all classes by minimizing this entropy of all the error vectors between the training feature vectors and the weight vector. To this end, we have proposed prudent learning model that favors competition with both the worst performer and the best performer based on the HT. The conversion of Renyi entropy function into the possibilistic domain helps us generate Renyi sigmoid and Renyi energy features. These new features and classifier are implemented on two datasets: Finger-knuckle-print for the authentication of persons and defect classification in the fabrics. The experimental results vindicate the effectiveness of the proposed features, classifier, and the learning model.