The previous chapter discussed how ACT-R can model human cognition assuming that the person has a certain set of knowledge structures with a certain set of parameters. That chapter displayed some examples of the relatively high degree of precision in the predictions of ACT-R models. For purposes of expository simplification, the examples were rather small but later chapters will display more complex models. The success of ACT-R's performance models makes the topic of this chapter all the more compelling. How did that knowledge get in there in the first place? Performance models should be learnable.
Table 2.1 classified ACT-R's assumptions about learning into a 2x2 scheme that paralleled the 2x2 classification of the performance assumptions. One dimension is whether the assumptions are concerned with the acquisition of procedural knowledge or whether they are concerned with the acquisition of declarative knowledge. The other dimension is whether they are concerned with symbolic learning (the acquisition of the chunks and productions themselves) or whether they are concerned with subsymbolic learning (the acquisition of the parameters that govern the deployment of these knowledge elements). The four sections of this chapter will address the topics defined by the four cells of this classification.
Knowledge Compilation Examples
Kendler and Kendler's Experiment