# BINARY LOGISTIC REGRESSION ESTIMATION USING A QUADRATIC PENALTY.  Returns 
# the maximum penalized likelihood estimate for the parameters of a logistic
# regression model.  The inputs in the training cases are in X, which should 
# be a matrix (or convertable to a matrix) with one row per training case, 
# and one column per input variable.  The training targets are in y, which
# should be a vector of 0/1 values.  The coefficients (but not the intercept) 
# are penalized by lambda (default 0) times the sum of their squares.
#
# The first element of the parameter vector returned is the intercept; the 
# remaining elements are the coefficients of the inputs.  

lr.est = function (y, X, lambda=0)
{
  X = as.matrix(X)  # Convert from data frame, if necessary

  result = nlm (function (beta) -lr.log.like(y,X,beta) + lambda*sum(beta[-1]^2),
                rep(0,ncol(X)+1), iterlim=200)

  beta = result$estimate
  attr(beta,"max.log.lik") = -result$minimum  # tack on max value as attr

  beta
}


# LOGISTIC REGRESSION LOG LIKELIHOOD.  Computes the log probability of the
# training data with inputs X and targets y, using the parameters in beta.

lr.log.like = function (y, X, beta)
{
  X = as.matrix(X)
  ys = 2*y-1

  lin = as.vector(beta[1] + X %*% beta[-1])

  sum (-log(1+exp(-lin*ys)))
}


# MAKE PREDICTIONS USING LOGISTIC REGRESSION COEFFICIENTS.  Returns the
# probability that y=1 for each of the test cases whose inputs are in X,
# using the intercept and coefficient parameters in beta.

lr.pred = function (X, beta)
{
  X = as.matrix(X)

  lin = as.vector (beta[1] + X %*% beta[-1])

  1/(1+exp(-lin))
}