Mining the space of generality with uncertainty-concerned cooperative classifiers

Generality is a recurrent theme in automated inductive systems. Induction of general patterns/rules is of course complicated by several factors. For example, higher levels of uncertainty and error are naturally introduced by generality. Moreover, it is not clear what sort of trade-off should be sought between increasing generality and decreasing predictive power. As a result, specific criteria to guide the search for useful general rules do not abound. In this paper, I reconsider these issues in the context of the generalized, fuzzy-like classifier system first proposed by Frey and Slate (1991) and later equipped with a Bayesian learning component by Muruzabal (1998). A crucial feature of this approach is that uncertainty is probabilistically measured at each classifier in the population. A new reinforcement policy exploiting this probabilistic structure and priming cooperation among general classifiers is introduced and shown to promote the stability of niches of reasonably high predictive power. The underlying genetic algorithm contributes effectively to learning although it somehow counteracts the built-in bias towards generality.