Lookahead Planning and Latent Learning in a Classifier System

Classifier systems (CSs) have been used to simulate and describe the behavior of adaptive organisms, animats, and robots. However, classifier system implementations to date have all been reactive systems, which use simple S-R rules and which base their learning algorithms on trial-and-error reinforcement techniques similar to the Hullian Law of Effect. While these systems have exhibited interesting behavior and good adaptive capacity, they cannot do other types of learning which require having explicit internal models of the external world, e.g., using complex plans as humans do, or doing “latent learning” of the type observed in rats. This paper describes a classifier system that is able to learn and use internal models both to greatly decrease the time to learn general sequential decision tasks and to enable the system to exhibit latent learning.