#
# The function loads the "reuters" sample data, necessary for the XIV exercise session.
# Note that it requires internet connection. 
# If you don't like it, download the reuters.txt file mentioned in line 19 manually,
# put it in the working directory and change the read.table call to refer to the local copy.
# Note that you can download the compressed version (reuters.txt.gz). In this case
# you do not need to unpack the gz - for local files R handles unpacking automagically.
#
load.data = function() {
  # The following is a slightly-processed version of the Reuters dataset. Namely:
  #   All articles with no TOPIC annotations are dropped
  #   The text of each article is converted to lowercase, whitespace is normalized to single-spaces.
  #   Only the first term from the TOPIC annotation list is retained (some articles have several topics assigned).
  #   The resulting dataset is a list of pairs (Topic, News content)
  #
  # This is a fun dataset to play with, so I suggest you play with it beyond what we do in the exercise session.
  #
  reuters = read.table("https://courses.cs.ut.ee/2014/ml/uploads/Main/reuters.txt", header=T)
  
  # We leave only two topics here: Crude Oil and Grain-related news.
  reuters = reuters[reuters$Topic == "crude" | reuters$Topic == "grain",]
  
  # In addition, to avoid waiting for too long, we limit our experiments to just 500 elements (300 train + 200 test).
  set.seed(1)
  reuters = reuters[sample(1:nrow(reuters), 500),]
  
  # Fixup the columns
  reuters$Content = as.character(reuters$Content) # R originally loads this as factor.
  reuters$Topic = factor(reuters$Topic) # Re-level the factor to have only two levels
  reuters$y = 2*(reuters$Topic == "grain") - 1
  reuters
}


#
# The Kernel Perceptron algorithm. Requires some fixing before use.
#
kernel_perceptron = function(K, y) {
  alpha = rep(0, nrow(K))
  b = 0
  while (TRUE) {
    # Make predictions for all points in the training sample
    predictions = 0 # <----- FIXME
    
    # Find misclassified instances
    misclassified = which(sign(predictions) != y)
    
    # If no misclassified, we're done!
    if (length(misclassified) == 0) { break; }
    
    # If something is misclassified, pick the first element
    i = misclassified[1]
    
    # Update parameters
    alpha[i] = alpha[i] + 0   # <------- FIXME {Hint: in the usual perceptron this could be w := w + y[i]*x[i]}
    b = b + 0                 # <------- FIXME {Hint: in the usual perceptron this could be b := b + y[i]}
  }
  
  # Return result
  result = list(alpha, b)
  names(result) = c("alpha", "b")
  result
}

circle.data = function() {
  angles = runif(300)*2*pi
  norms = c(runif(100), runif(100) + 2, runif(100) + 4)
  x = sin(angles)*norms
  y = cos(angles)*norms
  c = factor(c(rep(1,100), rep(2,100), rep(3,100)))
  data.frame(x=x, y=y, c=c)
}