Gluck and Bower (1988) adaptive least-mean-square (LMS) network
slpLMSnet(st, tr, xtdo = FALSE, dec = "logistic")
List of model parameters
Numerical matrix of training items, use
Boolean specifying whether to include extended information in the output (see below)
Specify what response rule to use.
The function operates as a stateful list processor (slp; see Wills et al., 2017). Specifically, it takes a matrix as an argument. Each row represents a single trial. Each column represents different types of information required by the implementation of the model, such as the elemental representation of stimuli, teaching signals, and other variables specifying the model's behaviour (e.g. freezing learning).
st must be a list containing the following items:
beta - the learning rate (fixed for a given simulation) for the
LMS learning rule. The upper bound of this parameter is not
specified, but we suggest 0 < beta ≤ 1.
theta - is a positive scaling constant. When theta rises, the
logistic choice function will become less linear. When theta is
high, the logistic function will approximate the behaviour of a step
bias - is a bias parameter. It is the value of the output
activation that results in an output probability rating of P =
0.5. For example, if you wish an output activation of 0.4 to produce a
rated probability of 0.5, set beta to 0.4. If you are not sure what to
use here, set it to 0. The bias parameter is not part of the original
Gluck and Bower (1988) LMS network, see Note 1.
w - is a matrix of initial connection weights, where each row is
an outcome, and each column is a feature or cue. If you are not sure
what to use here, set all values to 0.
outcomes - is the number of possible categories or outcomes.
colskip - the number of optional columns to be skipped in the tr
matrix. colskip should be set to the number of optional columns
PLUS ONE. So, if you have added no extra columns, colskip = 1.
tr must be a matrix, where each row is one trial
presented to the model. Trials are always presented in the order
specified. The columns must be as described below, in the order
ctrl - a vector of control codes. Available codes are: 0 = normal
trial; 1 = reset model (i.e. set associative strengths (weights)
back to their initial values as specified in w (see above)); 2 =
Freeze learning. Control codes are actioned before the trial is
opt1, opt2, ... - any number of preferred optional columns, the
names of which can be chosen by the user. It is important that these
columns are placed after the control column, and before the
remaining columns (see below). These optional columns are ignored by
the slpLMSnet function, but you may wish to use them for
readability. For example, you might choose to include columns such
as block number, trial number and condition. The argument colskip
(see above) must be set to the number of optional columns plus one.
x1, x2, ... - activation of input nodes of corresponding features.
Feature patterns usually represented as a bit array. Each element in the
bit array encodes the activations of the input nodes given the presence or
absence of the corresponding features. These activations can take on either
1 or 0, present and absent features respectively. For example, Medin and
Edelson's (1988) inverse base-rate effect with stimuli AB and AC can be
represented as [1 1 0] and [1 0 1] respectively. In a more unconventional
scenario, you can set activation to vary between present 1 and absent -1,
see Note 2. slpLMSnet can also support any positive or negative real number
for activations, e.g. one might use values between 0 and 1 to represent the
salience of the features.
d1, d2, ... - teaching input signals indicating the category feedback
on the current trial. It is a bit array, similar to the activations of
input nodes. If there are two categories and the stimuli on the current
trial belongs to the first, then this would be represented in
[1 0], on edge cases see Note 3. The length of this array must be provided
Returns a list with the following items if
xtdo = FALSE:
A matrix with either the probability rating for each
outcome on each trial if
Output node activations on each trial, as output by Equation 3 in Gluck and Bower (1988).
A connection weight matrix, W, where
each row represents the corresponding element in the teaching
signals array in
xtdo = TRUE, the following item also returned:
The least mean squeared differences between desired and actual activations of output nodes on each trial (Eq. 4 in Gluck and Bower, 1988). This metric is an indicator of the network's performance, which is measured by its accuracy.
bias parameter is not part of the original Gluck and
Bower (1988) model.
bias in the current implementation helps
comparisons between simulations using the
logistic choice function. Set bias to 0 for operation as specified
in Gluck & Bower (1988). Also note that, where there is more than
one output node, the same bias value is subtracted from the output
of each node. This form of decision mechanism is not present in the
literature as far as we are aware, although using a negative bias
value would, in multi-outcome cases, approximate a 'background
noise' decision rule, as used in, for example, Nosofsky et
2. slpLMSnet can support both positive and negative real numbers as input node activations. For example, one might wish to follow Markman's (1989) suggestion that the absence of a feature element is encoded as -1 instead of 0.
3. slpLMSnet can process a bit array of teaching signals, where the model is told that the stimulus belongs to more than one category. slpLMSnet uses matrix operations to update weights, so it can encode and update multiple teaching signals on the same trial.
Lenard Dome, Andy Wills
Gluck, M. A., & Bower, G. H. (1988). From conditioning to category learning: An adaptive network model. Journal of Experimental Psychology: General, 117, 227-247.
Markman, A. B. (1989). LMS rules and the inverse base-rate effect: Comment on Gluck and Bower (1988). Journal of Experimental Psychology: General, 118, 417-421.
Medin, D. L., & Edelson, S. M. (1988). Problem structure and the use of base-rate information from experience. Journal of Experimental Psychology: General, 117, 68-85.
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 Shepard, Hovland, and Jenkins (1961). Memory and Cognition, 22, 352-369.
Wills, A.J., O'Connell, G., Edmunds, C.E.R., & Inkster, A.B.(2017). Progress in modeling through distributed collaboration: Concepts, tools, and category-learning examples. Psychology of Learning and Motivation, 66, 79-115.
## load catlearn library(catlearn) ## create st with initial state st <- list(beta = 0.025, # learning rate theta = 1, # decision scaling parameter bias = 0, # decision bias parameter # initial weight matrix, # row = number of categories, # col = number of cues w = matrix(rep(0, 6*4), nrow = 4, ncol = 6), outcomes = 4, # number of possible outcomes colskip = 3) ## create inverse base-rate effect tr for 1 subject and without bias cue tr <- krus96train(subjs = 1, ctxt = FALSE) # run simulation and store output out <- slpLMSnet(st, data.matrix(tr)) out$connectionWeightMatrix
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.