Concept learning using complexity regularization

We apply the method of complexity regularization to learn concepts from large concept classes. The method is shown to automatically find a good balance between the approximation error and the estimation error. In particular, the error probability of the obtained classifier is shown to decrease as O(dm) to the achievable optimum, for large nonparametric classes of distributions, as the sample size n grows. We also show that if the Bayes error probability is zero and the Bayes rule is in a known family of decision rules, the error probability is O(log n/n) for many large families, possibly with infinite VC dimension.