Performance
John R. Anderson
Christian Lebiere
Marsha Lovett
Abstract
ACT-R aspires to provide detailed and accurate models of the learning and performance of subjects across a wide range of experimental paradigms. In cognitive psychology we typically measure the performance of subjects in two ways -- what responses they choose to emit and how long it takes them to emit these responses. The former measure often is presented as "percent correct" or its complement "percent errors". One can be judgmental in this binary way in cases where there is a designated correct response. However, in other cases there can be a richer category of potential responses and we simply calculate the percentage of responses that fill various categories. Usually, latency is measured as time to complete a task but in some cases we can collect latency measures for intermediate responses (e.g., eye movements, mouse movements, key presses) on the way to task completion. As discussed in the first chapter, production systems address these two dependent measures of psychology more directly and more completely than any other modeling formalism. Current production systems are committed to the exact timing and identity of each response a subject emits.
The behavior of the ACT-R system and, in particular, its predictions about these two dependent measures are a function of the productions that fire. This chapter will be examining which productions ACT-R chooses to fire, how it instantiates these productions, and what determines the latencies of these production firings. This performance analysis will assume that we have a static system that is not changing as a function of experience. This is a typical assumption in much of the experimental research on human cognition. It is justified in cases where the behavior under study is at some relatively asymptotic level or the critical factors being investigated do not change over the range of experiences encountered in the experiment. The next chapter will investigate learning issues -- how an ACT-R model changes itself as a function of its experience. However, even in this chapter much of the discussion will be motivated by implicit assumptions about the learning problem and at some points we will be forced into an explicit discussion of learning issues. The separation of performance from learning is basically an expositional strategy. They are quite intertwined in the ACT-R theory.
Models
Figure 3.1 Demo
Fan Effect Model
Probability Matching Models
Runquist Experiment Model
Siegler Experiment Model