edl (version 1.1)

updateWeightsNoOutcomeCompetition: Function implementing the Rescorla-Wagner learning equations without outcome competition (for illustration purposes) for a single learning event.

Description

Function implementing the Rescorla-Wagner learning equations without outcome competition (for illustration purposes) for a single learning event. A set of cues and outcomes are provided, and a weightmatrix that needs to be updated.

Usage

updateWeightsNoOutcomeCompetition(
  cur.cues,
  cur.outcomes,
  wm = NULL,
  eta = 0.01,
  lambda = 1,
  alpha = 0.1,
  beta1 = 0.1,
  beta2 = 0.1
)

Arguments

cur.cues

A vector with cues.

cur.outcomes

A vector with outcomes.

wm

A weightmatrix of class matrix. If not provided a new weightmatrix is returned. Note that the cues and outcomes do not necessarily need to be available as cues and outcomes in the weightmatrix: if not present, they will be added.

eta

Learning parameter, typically set to 0.01. If eta is not specified and set to the value NULL, the values of alpha, beta1, and beta2 determine the learning rate. However, changing these settings is generally not very useful (see Hoppe et al, submitted).

lambda

Constant constraining the connection strength.

alpha

Learning parameter (scaling both positive and negative evidence adjustments), typically set to 0.1.

beta1

Learning parameter for positive evidence, typically set to 0.1.

beta2

Learning parameter for negative evidence, typically set to 0.1.

Value

A weightmatrix (matrix)

See Also

RescorlaWagner, RWlearning

Other functions for explaining error-driven learning: RWlearningNoCueCompetition(), RWlearningNoOutcomeCompetition(), updateWeightsNoCueCompetition()

Examples

# NOT RUN {
# load example data:
data(dat)

# add obligatory columns Cues, Outcomes, and Frequency:
dat$Cues <- paste("BG", dat$Shape, dat$Color, sep="_")
dat$Outcomes <- dat$Category
dat$Frequency <- dat$Frequency1
head(dat)
dim(dat)

# now use createTrainingData to sample from the specified frequencies: 
train <- createTrainingData(dat)

# this training data can actually be used train network:
wm <- RWlearningNoOutcomeCompetition(train)
# retrieve trained network:
new <- getWM(wm)

train2 <- createTrainingData(dat)
updateWeightsNoOutcomeCompetition(getValues(train2$Cues[1]), 
    getValues(train2$Outcomes[1]), wm=new)

# comparison between eta and alpha, beta1, beta2:
check.cues <- c("BG", "car", "red")
new[check.cues,]
tmp1 <- updateWeightsNoOutcomeCompetition(check.cues, 
    c("vehicle", "animal"), wm=new)
tmp2 <- updateWeightsNoOutcomeCompetition(check.cues, 
    c("vehicle", "animal"), wm=new, eta=NULL)
tmp3 <- updateWeightsNoOutcomeCompetition(check.cues, 
    c("vehicle", "animal"), wm=new, beta1=0.2)
tmp4 <- updateWeightsNoOutcomeCompetition(check.cues, 
    c("vehicle", "animal"), wm=new, eta=NULL, beta1=0.2)
# these two should be the same:
tmp1[check.cues,]
tmp2[check.cues,]
# now we change beta2, but this does not change anything,
# because eta is being used:
tmp3[check.cues,]
# when we turn eta off, beta2 changes the values:
tmp4[check.cues,]

# }