However, it is also possible that the participants did not
individually categorize the stimuli. As their name implies,
the guessing models assumed that participants randomly
chose a response on each trial, without considering the
individual category membership of the stimuli. Finally, the
similarity model defined similarity as an exponentially
decreasing function of the weighted Euclidean distance and
assumed that participants responded ‘‘Same’’ for high
similarities and ‘‘Different’’ for low similarities. The similarity
model had two free parameters, one to differentially
weight the stimulus dimensions and another to describe the
rate of the exponential decrease. Like the guessing model,
the similarity model also does not assume separate classification
of the stimuli.
The results from the model-based analyses are shown in
Table 1. As can be seen, most participants in the rule-based
conditions appeared to be responding optimally. Furthermore,
note that the responses of these participants were
more likely to be best fit by an optimal model later in
training. A one-tail binomial test showed that the difference
in proportion of best-fitting optimal models between
early and late performance was statistically significant for
the 1D-Width condition (p.05), but not for the 1DOrientation
condition. Even so, the performance of most
participants in both rule conditions was best fit by an
optimal categorization model by the end of training. In the
information-integration condition, only one participant
used an optimal strategy at the beginning of the experiment,
and no participant used an optimal categorization
strategy by the end of the experiment. A one-tail binomial
test showed that this decrease in the proportion of optimal
best-fitting models was not significant. It should be noted
that none of the participants in any block was best fit by a
similarity model. Participants not best fit by an optimal
categorization model were best fit by a guessing model.