MACS (Modular Anticipatory Classifier System) is a new Anticipatory Classifier System. With respect to its predecessors, ACS, ACS2 and YACS, the latent learning process in MACS is able to take advantage of new regularities. Instead of anticipating all attributes of the perceived situations in the same classifier, MACS only anticipates one attribute per classifier. In this paper we describe how the model of the environment represented by the classifiers can be used to perform active exploration and how this exploration policy is aggregated with the exploitation policy. The architecture is validated experimentally. Then we draw more general principles from the architectural choices giving rise to MACS. We show that building a model of the environment can be seen as a function approximation problem which can be solved with Anticipatory Classifier Systems such as MACS, but also with accuracy-based systems like XCS or XCSF, organized into a Dyna architecture.