A Multinomial prototype-based learning algorithm

Bulut, Ahmet Can
Recent studies in machine learning field proved that ideas which were once thought impractical are in fact tangible. Over the years, researchers have managed to develop learning systems which are able to interact with the environment and use experiences for adaptation to new conditions. Humanoid robots can now learn concepts such as nouns, adjectives and verbs, which is a big step for building human-like learners. Behind all these achievements, development of successful learning and classification techniques is one of the key factors. In this thesis, we propose a novel prototype-based learning method which uses the distributional properties of class dimensions. By dealing with the problem of feature dimensions' having multiple polarities, our algorithm can distinguish the dimensions which display unpredictable behaviors from the ones which are composed of multiple predictable patterns. We tested our algorithm on 8 different datasets and compared the results with 9 other algorithms including SVM and AdaBoost. Apart from being insensitive to the ordering of inputs, our method showed that it provides comparable performance in terms of accuracy rate, running time, learning curve and most importantly the ability to resolve multipolarity in dimensions.