We present a learning algorithm for nominal data. It builds a classifier by adding iteratively a simple patch function that modifies the current classifier. Its main advantage lies in the possibility to learn every patch function parameters optimally from the Bayesian point of view hence avoiding overtraining.