ABSTRACT
We describe how to extend the ACT-R production system to model human errors in the performance of a high-level cognitive task: to solve simple linear algebra problems while memorizing a digit span. Errors of omission are produced by introducing a cutoff on the latency of memory retrievals. If a memory chunk cannot gather enough activation to be retrieved before the threshold is reached, retrieval fails. Adding Gaussian noise to chunk activation produces a pattern quantitatively similar to subject errors. Errors of commission are introduced by allowing imperfect matching in the condition side of productions. The wrong memory chunk can be retrieved if its activation is large enough to allow it to overcome the mismatch penalty. This mechanism provides a qualitative and quantitative fit to subject errors. In conclusion, this paper demonstrates that human-like errors, sometimes thought of as the exclusive domain of connectionist models, can be successfully duplicated in production system models.
