Using_an_optimizer_to_get_probability_threshold.R

install.packages("MASS")
library(MASS)
data(Pima.tr)
data(Pima.te)
#############


## STEP 1: Logistic regression, predict diabetes yes or no ##
logistic_reg <- glm(type ~ ., data = Pima.tr, family = binomial) # basic model 
predict_logistic.tr <- predict(logistic_reg, type = "response")  # predicted probabilities (TRAINING SET)

## STEP 2: Create a function that evaluates the misclassification rate for the TRAINING SET, for any threshold
evaluate_fn <- function(threshold = NA)
{
  predicted_outcomes <- as.numeric(predict_logistic.tr > threshold)
  table_logistic <- table(Pima.tr$type, predicted_outcomes)
  
  error_rate_logistic <- sum(table_logistic[2:3])/sum(table_logistic)
  return(error_rate_logistic)
}

## STEP 3: Optimize for threshold within TRAINING SET using a hill-climbing algorithm
best_threshold <- optim(0.5, evaluate_fn, method = "Brent", 
                        lower = 0, upper = 1)$par
cat("\nThe best threshold is:", best_threshold, "\n")
cat("\nThe error rate at this threshold is:", evaluate_fn(best_threshold), "\n\n")

## OPTIONAL STEP 4:  Optimize for threshold within TRAINING SET using a genetic algorithm
install.packages("rgenoud") 
library(rgenoud)
best_threshold_genetic <- genoud(fn = evaluate_fn, nvars = 1, max = FALSE, print.level = 1, starting.values = best_threshold)$par
cat("\nThe best (genetic algorithm-derived) threshold is:", best_threshold_genetic, "\n\n")
cat("\nThe error rate at this (genetic-derived) threshold is:", evaluate_fn(best_threshold_genetic), "\n\n")