Create randomized training blocks for CIRP nosof94, in a format
suitable for the slpALCOVE or slpSUSTAIN models, and
other models that use the same input representation formats.
nosof94train(cond = 1, blocks = 16, absval = -1, subjs = 1, seed = 7624,
missing = 'geo', blkstyle = 'accurate')Category structure type (1-6), as defined by Shepard et al. (1961).
Number of blocks to generate. Omit this argument to get the same number of blocks (16) as used in the simulations reported by Nosofsky et al. (1994).
Teaching value to be used where category is absent.
Number of simulated subjects to be run.
Sets the random seed.
If set to 'geo', output missing dimension flags (see
below). If set to 'pad', use the padded stimulus representation format
of slpSUSTAIN. If set to 'pad', set absval to zero.
If set to 'accurate', reproduce the randomization of this experiment, as described in Nosofsky et al. (1994). If set to 'eights', use instead the randomization used in the Gureckis (2016) simulation of this experiment.
R by C matrix, where each row is one trial, and the columns contain model input.
A matrix is produced, with one row for each trial, and with the following columns:
ctrl - Set to 1 (reset model) for trial 1 of each simulated
subject, set to zero (normal trial) for all other trials.
blk - training block
stim - Stimulus number, ranging from 1 to 8. The numbering scheme
is the same as in Nosofsky et al. (1994, Figure 1), under the mapping
of dim_1_left = 0, dim_1_right = 1, dim_2_front = 0, dim_2_back = 1,
dim_3_bottom = 0, dim_3_top = 1.
x1, x2, … - input representation. Where missing='geo',
x1, x2, and x3 are returned, each set at 1 or 0. This is the binary
dimensional representation required by models such as slpALCOVE, where
e.g. x2 is the value on the second dimension. Where
missing='pad', x1, x2, y1, y2, z1, z2, are returned. This is the
padded represenation required by models such as slpSUSTAIN; e.g. y1 and
y2 represent the two possible values on dimension 2, so if y1 is black,
y2 is white, and the stimulus is white, then [y1, y2] = [0, 1].
t1, t2 - Category label (1 = category present, absval = category
absent)
m1, m2, m3 - Missing dimension flags (always set to zero in this
experiment, indicating all input dimensions are present on all
trials). Only produced if missing = 'geo'.
Although the trial ordering is random, a random seed is used, so multiple calls of this function with the same parameters should produce the same output. This is usually desirable for reproducibility and stability of non-linear optimization. To get a different order, use the seed argument to set a different seed.
This routine was originally developed to support Wills et al. (n.d.).
Nosofsky, R.M., Gluck, M.A., Plameri, T.J., McKinley, S.C. and Glauthier, P. (1994). Comparing models of rule-based classification learning: A replication and extension of Shepaard, Hovland, and Jenkins (1961). Memory and Cognition, 22, 352--369
Gureckis, T. (2016). https://github.com/NYUCCL/sustain_python
Wills et al. (n.d.). Benchmarks for category learning. Manuscript in preparation.