How do people solve the "weather prediction" task?: individual variability in strategies for probabilistic category learning

Abstract

Probabilistic category learning is often assumed to be an incrementally learned cognitive skill, dependent on nondeclarative memory systems. One paradigm in particular, the weather prediction task, has been used in over half a dozen neuropsychological and neuroimaging studies to date. Because of the growing interest in using this task and others like it as behavioral tools for studying the cognitive neuroscience of cognitive skill learning, it becomes especially important to understand how subjects solve this kind of task and whether all subjects learn it in the same way. We present here new experimental and theoretical analyses of the weather prediction task that indicate that there are at least three different strategies that describe how subjects learn this task. (1) An optimal multi-cue strategy, in which they respond to each pattern on the basis of associations of all four cues with each outcome; (2) a one-cue strategy, in which they respond on the basis of presence or absence of a single cue, disregarding all other cues; or (3) a singleton strategy, in which they learn only about the four patterns that have only one cue present and all others absent. This variability in how subjects approach this task may have important implications for interpreting how different brain regions are involved in probabilistic category learning.

Figure 1

The four cards used as cues in Knowlton et al. (1994). Each card was associated with each possible outcome with a fixed probability.

Figure 2

Percent optimal responses (A) over all 200 trials, (B) over first 50 trials of training.

Figure 3

Actual and estimated cue probabilities.

Figure 4

Percentage of subjects best fit by various strategy models across the four training blocks.

Figure 5

Percent optimal responding as a function of best-fit model.

Figure 6

Percent optimal responses (A) over all 200 trials, (B) over first 50 trials of training.

Figure 7

The percentage of subjects for whom multi-cue, one-cue, and singleton strategies provided a best-fit model for each of the four training blocks.

Figure 8

Percent optimal responding as a function of best-fit strategy.