In a recent study, verbal protocols were used to investigate the nature of changes in low-level interactions that take place during learning with a computerized tutor called Stat Lady (Shute & Gluck, 1994). Analyses showed consistent behavioral changes in the use of the tutor interface, which account for a substantial portion of the learning curve, independent of error rates. In fact, more than half of the learning curve could be accounted for by these changes in low-level interactions. The changes primarily are decreases in the verbalization of on-screen text, although the elimination of interface confusion also contributes to the efficiency gain.
We characterize these changes as resulting from (1) the acquisition of knowledge regarding the sequence of activities in the scenarios - which allows for less reading of instruction text and more recall from memory, and (2) adaptive changes in reading behaviors, such that redundant and uninformative text comes to be ignored more often.
However, demonstrating that these are reasonable explanatory mechanisms requires a running model that can reproduce the results. We have used ACT-R (Anderson & Lebiere, 1998), to develop a production system model that not only matches the verbal protocol and latency data from Experiment 1, but does so using the mechanisms we have postulated above. It serves as a proof-of-concept for these interface learning mechanisms, and also as a demonstration of the applicability of the ACT-R architecture for increasing our understanding of phenomena occurring at the interface of the human and the computer.