Complexity regularization using data-dependent penalties

We define a regression function estimate based on complexity regularization, where the list of candidate functions and the corresponding penalties are determined from the training data, leading to improved performance.