Prior to the actual process of model induction, the learning problem needs to be formalized and modeled appropri-ately; typically, this includes the specification of various data spaces, the mathematical structure of these spaces, the relationship between them, etc. Successful learning requires a suitable formalization of the problem, which is a point that is often overlooked in machine learning. Fuzzy logic has much to offer in this regard, and definitely more than what has been realized so far.
First, one should note that “fuzzy modeling” is not restricted to expressing functional dependencies (fuzzy rules, fuzzy decision trees). Instead, there is much more that can be modeled and formalized in terms of fuzzy concepts (aggregation functions, similarity relations, etc.), albeit in a more subtle way. For example, we already mentioned the idea of “modeling data” [17]. Likewise, the structure of underlying data spaces can be characterized in terms of fuzzy relations, for example by equipping them with fuzzy order relations [7,29], on which learning algorithms can then operate conveniently.
Second, modeling becomes even more an issue in settings that go beyond standard supervised learning, such as constructive induction or reinforcement learning—in the latter, for example, a suitable abstraction of the state space is of critical importance.