MainLoader.r

getwd()
setwd("/Users/ian/dev/ML/week-three")
data <- read.csv("dataset.csv")

# load perceptron algorithm
source("Perceptron.R")

# model graph the data
perceptron.model <- perceptron(data)
perceptron.model(c(2,2))

peceptron.r

# Perceptron calculates the decision surface depending on the 
# weights of the feature vectors

perceptron <- function (data)
{
	data.matrix <- as.matrix(data)

	df <- subset(data, select = x1: x2)
	mat <- as.matrix(df)
	counter <- 0

	# learner rate
	n <- 0.3

	# initial bias
	#b <- 0
	#b <- -2
	b <- c(sample(-3:3,1)
	
	# initial weight vector
	#w <- c(0,0)
	#w <- c(0,2)
	w <- c(sample(-1:1,1),sample(-1:1,1))
	
	# maximum radius of the furthest vector (r)
	norm <- mat^2
	row.sum <- max(rowSums(norm, na.rm = FALSE, dims = 1))
	r <- sqrt(row.sum)

	data.matrix[4, ncol(data.matrix)]
	sign(w %*% mat[4,] - b)

	# the perceptron algorithm
	repeat
	{
		failed <- FALSE
		for (i in 1: nrow(data.matrix))
		{
			yi <- data.matrix[i, ncol(data.matrix)]
			if (sign(w %*% mat[i,] - b) != yi)
			{

			  counter <- counter + 1
			    
				# calculate the new weights (w) 
			  w <- w + n * mat[i,] * yi
			
				# calculate the new bias (b) value
			  b <- b - n * yi * r^2

			  cat("Iteration", counter, "\n")
				print(b)
				
			}
		}

		for (i in 1: nrow(data.matrix))
		{
			print((sign(w %*% mat[i,] - b)))

			# check if the signs match
			if (sign(w %*% mat[i,] - b) != data.matrix[i, ncol(data.matrix)])
			{
				failed <- TRUE
			}
		}
		# if all signs pass, learning is complete
		if(!failed)
			break   
	}

	# the decision surface is perpendicular to w
	slope <- -w[1]/w[2]  # d = w = (w1, w2), slope of w = w[2]/w[1]
	intercept <- b/w[2]

	# Seperate the positive data points from the negative ones
	pos <- data[which(data[,3]>0), -3]

	# Plot the data points
	plot(data$x1, data$x2, col = "green", pch = 19, xlab = "X1", ylab = "X2")

	# Add the positive points back to the plot
	points(pos$x1, pos$x2, col="blue", pch=19)

	# Draw the decision surface
	abline(intercept, slope, col = "red" )

	# perceptron prediction function
	function(x)
	{
		val <- (w %*%x - b)
		sign(val)
	}
}