Supervisory control of declarative and procedural responding

Abstract There is overwhelming evidence that multiple systems mediate human learning. If so, then a problem of fundamental importance is to determine how control is passed back and forth among these systems. Although more research is needed, existing evidence suggests that system switching does occur, but is much less common than existing models assume, and that the bias toward strategies mediated by declarative memory is stronger than previously assumed. These shortcomings are addressed by testing a variety of different alternative models of system switching against several empirical benchmarks. The models all assumed separate explicit-rule and procedural/exemplar-similarity categorization systems. The only models that passed all benchmark tests incorporated a switch cost and switched systems when explicit-system accuracy dropped below a criterion value and procedural system accuracy was simultaneously reasonably high. This research suggests a new class of system-switching models, which could be incorporated into any existing multiple-systems model (e.g., COVIS or ATRIUM), and that is much more consistent with existing data than current models.