Noise modelling and evaluating learning from examples

AbstractThe means of evaluating, using artificial data, algorithms, such as ID3, which learn concepts from examples is enhanced and referred to as the method of artificial universes. The central notions are that of a class model and its associated representations in which a class attribute is treated as a dependent variable with description attributes functioning as the independent variables. The nature of noise in the model is discussed and modelled using information-theoretic ideas especially that of majorisation. The notion of an irrelevant attribute is also considered. The ideas are illustrated through the construction of a small universe which is then altered to increase noise. Learning curves for ID3 used on data generated from these universes are estimated from trials. These show that increasing noise has a detrimental effect on learning