Rubyでパーセプトロン （１次元２クラスにしか対応してない版）

main.rb

require "./perceptron"
require 'gnuplot'

# クラス1とクラス2それぞれの学習パターン
learning_patterns = [
  [12, 5, 2, -2],
  [-4, -6, -15]
]

perceptron = Perceptron.new(learning_patterns)
perceptron.learn

#
# 特徴空間をプロット
#
output_file = "./feature_space.png"
Gnuplot.open do |gp|
  Gnuplot::Plot.new( gp ) do |plot|

    plot.title "feature space"
    plot.set "terminal png size 640, 300"
    plot.output output_file
    plot.xrange "[-20:20]"
    plot.yrange "[-1:1]"
    plot.xlabel "x"

    # 数直線
    lx = [*-20..20].map(&:to_f)
    ly = lx.map { |_| 0 }
    plot.data << Gnuplot::DataSet.new( [lx, ly] ) do |ds|
      ds.with = "lines"
      ds.linecolor = "rgb 'black'"
      ds.notitle
    end

    # 各学習パターン
    learning_patterns.each_with_index do |patterns, i|
      px = patterns
      py = px.map { |_| 0 }
      plot.data << Gnuplot::DataSet.new( [px, py] ) do |ds|
        ds.with = "points ls 1 pt 7 ps 2 lc #{i}"
        ds.notitle
      end
    end

    # 決定境界
    y = (-1..1).map(&:to_f)
    x = y.map do |_|
      -perceptron.weight_vector[0].to_f /
      perceptron.weight_vector[1].to_f
    end
    plot.data << Gnuplot::DataSet.new( [x, y] ) do |ds|
      ds.with = "lines ls 1 lc 2 lw 2"
      ds.notitle
    end
  end
end

#
# 重み空間をプロット
#
output_file = "./weight_space.png"
Gnuplot.open do |gp|
  Gnuplot::Plot.new( gp ) do |plot|

    plot.title  "weight space"
    plot.set "terminal png size 400, 420"
    plot.output output_file
    plot.xrange "[-3:6]"
    plot.yrange "[-3:6]"
    plot.xlabel "W1"
    plot.ylabel "W0"

    # 各学習パターンに対応した超平面 (赤色)
    learning_patterns.flatten.each do |pattern|
      x = [*-20..20].map(&:to_f)
      y = x.map { |v| -pattern * v }
      plot.data << Gnuplot::DataSet.new( [x, y] ) do |ds|
        ds.with = "lines ls 1 lc rgb 'red'"
        ds.notitle
      end
    end

    # 重みベクトルの更新履歴 (青色)
    vx = perceptron.log.map(&:rotate)
    vy = vx.map { |_| 0 }
    plot.data << Gnuplot::DataSet.new( [vx, vy] ) do |ds|
      ds.with = "linesp ls 1 pt 7 lw 2 lc rgb 'blue'"
      ds.notitle
    end
  end
end

perceptron.rb

class Perceptron

  RHO = 1
  attr_accessor :weight_vector, :log

  def initialize( learning_patterns )
    @learning_patterns = learning_patterns.map do |patterns|
      patterns.map { |pattern| [1, pattern] }
    end
    @weight_vector = [1, 1]
    @log = []
  end

  # 識別関数 g(x)
  def discriminate( pattern )
    @weight_vector.zip( pattern ).inject(0) do |sum, wx|
      sum + wx.inject(:*)
    end
  end

  # 誤識別があった場合重みベクトルを修正する
  def correct_errors
    @learning_patterns.size.times do |i|
      @learning_patterns[i].each do |pattern|
        case i
        # w' = w + ρx
        when ->(c) { c == 0 && discriminate( pattern ) <= 0 }
          @log << @weight_vector
          @weight_vector =
            @weight_vector.zip( pattern.map { |e| RHO * e } )
            .map { |a| a.inject(:+) }

        # w' = w - ρx
        when ->(c) { c == 1 && discriminate(pattern) >= 0 }
          @log << @weight_vector
          @weight_vector =
            @weight_vector.zip( pattern.map { |e| RHO * e } )
            .map { |a| a.inject(:-) }
        end
      end
    end
  end

  # 重みベクトルが更新されなくなるまで誤り訂正を繰り返す
  def learn
    prev = nil
    while prev != @weight_vector
      prev = @weight_vector
      correct_errors
    end
    @log << @weight_vector
  end
end