We study the problem of estimating some unknown regression or autoregression function in a B-mixing dependent framework (for the design or the errors). For this end, we consider some collection of models which are finite dimensional spaces. A penalized least-squares estimator (PLSE) is built on...