# パーセプトロン(2クラスかつ線形分類可能なデータのみ)

perceptron.rb

# coding: utf-8

require "pp"

# dimention次元の特徴に対する拡張重みベクトルを
# 乱数によって初期化
def initialize_weight(dimention)
  weight = Array.new(dimention + 1)
  (dimention + 1).times do |i|
    weight[i] = rand()
  end

  weight
end

# 内積
def inner_product(left, right)
  dimention = left.size

  sum = 0;
  dimention.times do |i|
    sum = sum + left[i] * right[i]
  end

  sum
end

# 重みを更新する量
def update_value(rho, feature_vector)
  value = Array.new(rho.size)

  rho.size.times do |i|
    value[i] = rho[i] * feature_vector[i]
  end

  value
end

# 識別関数 >= 0 ならばc(x) = 0
# 識別関数 <  0 ならばc(x) = 1
# として、重みを学習
def learn_perceptron(train_data, train_class, rho)
  dimention = train_data[0].size
  weight    = initialize_weight(dimention)
  rho = Array.new(dimention + 1, rho)

  # 全ての訓練データを正しく識別できる重みが見つかるまで繰り返す
  begin
    is_updated = false
    train_data.each_with_index do |t, i|
      # 拡張特徴ベクトル
      expand_feature = [1] + t
      # 識別関数
      discriminant = inner_product(weight, expand_feature)

      # w = w + rho * x   if c(x) = 0 && predict(x) == 1
      # w = w - rho * x   if c(x) = 1 && predict(x) == 0
      if discriminant >= 0 && train_class[i] == 1
        update_value = update_value(rho, expand_feature)
        (dimention + 1).times do |i|
          weight[i] = weight[i] - update_value[i]
        end
        is_updated = true
      elsif discriminant < 0 && train_class[i] == 0
        update_value = update_value(rho, expand_feature)
        (dimention + 1).times do |i|
          weight[i] = weight[i] + update_value[i]
        end
        is_updated = true
      else
      end
    end
  end while is_updated

  weight
end

def classfier(data, weight)
  dimention = data[0].size

  data.each do |t|
    expand_feature = [1] + t
    discriminant = inner_product(weight, expand_feature)
    if discriminant > 0
      print "0, "
      pp t
    else
      print "1, "
      pp t
    end
  end
end

train_data = []
train_class = []
50.times do |i|
  a = rand(-5.0..5.0)
  b = rand(-5.0..5.0)
  train_data << [a, b]

  if b >= a
    train_class << 0
  else
    train_class << 1
  end
end

weight = learn_perceptron(train_data, train_class, 2.0)

test = []
20.times do |i|
  test << [rand(-5..5), rand(-5..5)]
end
classfier(test, weight)