| |
Abstract:
Nosofsky et al. (1996) have demonstrated the superiority of
the ALCOVE (Kruschke, 1992) model on an impressive number of tasks.
The model's clear superiority in fitting human data, however, is
often held in favour of a cluster of theoretical assumptions that
the model is said to represent. For instance, the model assumes
euclidean distance based on Shepard's seminal work, a generalized
context model, an exemplar representation (in the form of Radial
Basis Functions), attentional gating of inputs etc. Some of these
properties have been tested against other models, for instance,
Nosofsky et al. have attempted to fit data with Gluck and Bower's
configural cue model and attentional weights. However, it is not
clear which of the many properties that the models differ on are
responsible for the successful fitting to human data. We propose to
list some of the theoretical properties that contribute to the
curve fitting and modify them in alcove to determine whether they
are essential or accidental in fitting the data. Specifically, we
look at: (1) representational model (exemplars) (2) the metric used
(euclidean) (3) the probability function (generalized context) (4)
the similarity function (exponential) and attempt to assess which
are important in the performance of the model. Preliminary
replication of Kruschke's simulations confirm that Shepard's task
is fit appropriately by the model and that attention is crucially
responsible for the appropriate behaviour.
|
