A new procedure for detecting departures from constancy of the parameters of a probabilistic model over a period of time is proposed. It basically consists of specification of a joint probability distribution for data and models and minimization of the total gain in information. In contrast to conventional techniques, this procedure can be used to determine both the number of different models and the positions of the points at which the switches from one model to another occur. It is demonstrated how this new method can be applied to the problem of identification of different stages of EEG background activity. The lengths of these stages might serve as indicators for drug effects in the EEG.