gputools (version 0.1-1)

gpuLr: A Toy Logistic Regression Routine

Description

This function is an demonstration of logistic regression. The matrix QR decomposition implemented by this function is not quite up to snuff. The regression is not numerically stable and doesn't handle inputs that give rise to rank deficient matrices. We hope to release a better version of this in the future.

Usage

gpuLr(obs, outcomes, epsilon=0.00001, ridge=0.00001, maxiter=200)

Arguments

obs

a matrix of floating point numbers. Each column is an observed set of parameters (inputs) for the logistic function.

outcomes

a vector of floating point numbers. Each value is an observed output when the corresponding column of the 'obs' argument is input into the logistic function.

epsilon

a floating point number. When the sum of squares of the difference between outputs of the logistic functions on input of obs is less than the number of columns of obs times epsilon squared, the regression is declared to have converged and the i

ridge

a floating point number. Amount of error to introduce during the iterations of the regression to aid convergence.

maxiter

an integer. Maximum number of iterations that the regression will be allowed to execute without converging.

Value

• a vector of floating point numbers of length 'nrow(obs)'. The coefficients of the logistic function fitted to 'obs' and 'outcomes'.

See Also

glm

Examples

numParameters <- 5
numExperiments <- 5
observationMat <- matrix(runif(numParameters*numExperiments), numParameters, 
	numExperiments)
obsOutcomes <- runif(numExperiments)
for(i in 1:numExperiments) {
	if(obsOutcomes[i] > 0.5) {
		obsOutcomes[i] <- 1
	} else {
		obsOutcomes[i] <- 0
	}
}
gpuLr(observationMat, obsOutcomes)