This paper follows on from earlier work that our colleagues and ourselves presented at IJCNN 2001, IJCNN 2002 and FUZZ-IEEE 2002. We referred to simulations of a recurrent network and of an adaptive system that was partly based on a recurrent network. Both models were successful in simulating human sequence learning in a reaction time paradigm that is widely used in cognitive science and experimental psychology. We argued that these models were not only successful in simulating human learning, but also in predicting successful generalization to novel sequences where humans show generalization, and a failure to generalize to novel sequences when humans fail. In this paper we present data from a novel experiment and novel simulations on a longer version of this serial reaction time task. Under these conditions the task appears to be more difficult to master for the human subjects and therefore more complex. Accordingly, humans were not able to learn that task at all. Their failure is predicted by both computational models. Combining these results with the earlier findings from the previous conferences suggests that both successful and unsuccessful human performance can be predicted by the computational models considered here.

It has been shown that humans can rely on both rules or associations to solve a problem. We present a model in which rules may be applied to a particular sequence learning task, but rather than the rules being applied in an all-or-none fashion, a continuum from fully representing to not representing a rule is required in order to model human task performance.