deepbeige 0.2.0 → 0.2.1

data/VERSION

@@ -1 +1 @@
-0.2.0
+0.2.1

data/deep_beige.rb

@@ -75,40 +75,15 @@ class DeepBeige
       puts "Evolving Generation #{generation_number}"
       player_number = 0
       @population.each do |neuralnet|
-        player1 = DeepBeige.new 
-        player1.neural_net = neuralnet
-        player1.game_name = @game_name
-        
-        5.times do
-          game = game.class.new
-          game.quiet = true
-          opponent_number = rand(@population.count)
-          #puts "#{player_number} versus opponent #{opponent_number}"
-          opponent_net = @population[opponent_number]
-          player2 = DeepBeige.new
-          player2.neural_net = opponent_net
-          player2.game_name = @game_name
-          
-          players = [player1,player2]
-          table = Table.new game, players
-          table.quiet = true
-          table.play_game
-          if game.drawn?
-            players.each do |player|
-              scores[player.id] +=1
-            end
-
-          elsif game.won?
-            winner = players[game.winner]
-            players.each do |player|
-              if player.id == winner.id
-                scores[player.id] +=2
-              else
-                scores[player.id] -=2
-              end
-            end
-          end
+        player = DeepBeige.new 
+        player.neural_net = neuralnet
+        player.game_name = game.name
 
+        5.times do
+          play_as_player game.name, player, 1, scores
+        end
+        5.times do 
+          play_as_player game.name, player, 2, scores
         end
         player_number += 1
       end 
@@ -208,6 +183,41 @@ private
     end
     save_population name
   end
+  
+private
+  def play_as_player game_name, player, player_number, scores
+    game = game_from_name game_name
+    game.quiet = true
+    opponent_number = rand(@population.count)
+    #puts "#{player_number} versus opponent #{opponent_number}"
+    opponent_net = @population[opponent_number]
+    opponent = DeepBeige.new
+    opponent.neural_net = opponent_net
+    opponent.game_name = @game_name
+    players = [player, opponent]
+    if player_number == 2
+      player = [opponent,player]
+    end
+
+    table = Table.new game, players
+    table.quiet = true
+    table.play_game
+    if game.drawn?
+      players.each do |player|
+        scores[player.id] +=1
+      end
+
+    elsif game.won?
+      winner = players[game.winner]
+      players.each do |player|
+        if player.id == winner.id
+          scores[player.id] +=2
+        else
+          scores[player.id] -=2
+        end
+      end
+    end
+  end
 end
 
 

data/deepbeige.gemspec

@@ -5,11 +5,11 @@
 
 Gem::Specification.new do |s|
   s.name = %q{deepbeige}
-  s.version = "0.2.0"
+  s.version = "0.2.1"
 
   s.required_rubygems_version = Gem::Requirement.new(">= 0") if s.respond_to? :required_rubygems_version=
   s.authors = ["David Bochenski"]
-  s.date = %q{2010-09-04}
+  s.date = %q{2010-09-05}
   s.description = %q{An AI learning program that plays board games}
   s.email = %q{david@bochenski.co.uk}
   s.extra_rdoc_files = [

data/main.rb

@@ -27,7 +27,11 @@ def player_vs_deepbeige db, game
   #ok so now I'm interested in playing my best creation
   db.start_game game.name
   me = Human.new game
-  play_game game, me, db, []
+  if which_player == "1"
+    play_game game, me, db, []
+  else
+    play_game game, db, me, []
+  end
 end
 
 def player_vs_player game
@@ -36,6 +40,12 @@ def player_vs_player game
   play_game game, p1,p2, []
 end
 
+def which_player
+  puts
+  puts "Would you like to be player 1 or 2?"
+  gets.chop
+end
+
 def options
   puts
   puts "What would you like to do?"

data/neural_net.rb

@@ -9,6 +9,7 @@ class NeuralNet
   def initialize
     @id = UUID.new.to_s.split(':')[1].chop
     @network = []
+    @sigma = 0.05
   end
     
   def evaluate 
@@ -46,7 +47,7 @@ class NeuralNet
     @network = []
     input_nodes = []
     inputs.times do 
-      input_nodes << Node.new
+      input_nodes << Node.new(@sigma)
     end
     if input_nodes.count > 0
       @network << input_nodes
@@ -54,31 +55,34 @@ class NeuralNet
     (tiers - 2).times do
       tier = []
       10.times do
-        tier << Node.new  
+        tier << Node.new(@sigma)
       end
       @network << tier
     end
     
     output_nodes = []
     outputs.times do
-      output_nodes << Node.new
+      output_nodes << Node.new(@sigma)
     end
     if output_nodes.count >0
       @network << output_nodes
     end
     link_tiers
+    recalculate_tau
   end
   
   def fingerprint
     topline = ""
     fingerprint = ""
+    sigma = ""
+    tau = ""
     @network.each do |tier|
       topline << "#{tier.count},"
       tier.each do |node| 
         fingerprint << node.fingerprint
       end
     end  
-    topline.chop + "\n" + fingerprint
+    topline.chop + "\n" + @sigma.to_s + "\n" + @tau.to_s + "\n" + fingerprint
   end
   
   def reload fingerprint
@@ -86,6 +90,10 @@ class NeuralNet
     i = 0
     tiers = fingerprint[i].split(',').to_a
     i += 1
+    @sigma = fingerprint[i].to_f
+    i += 1
+    @tau = fingerprint[i].to_f
+    i += 1
         
     @network = []
     tiers.each do |tier|
@@ -93,7 +101,7 @@ class NeuralNet
       tier.to_i.times do
         node_fingerprint = fingerprint[i]
         i += 1
-        node = Node.new
+        node = Node.new(@sigma)
         node.reload node_fingerprint
         nodes << node
       end
@@ -110,8 +118,12 @@ class NeuralNet
     #the ability to mutate the number of
     #nodes and their configuration
     #focussing instead on simple node weight mutation
+    
+    #first we mutate sigma
+    @sigma = @sigma * Math.exp(@tau * gaussian_random)
     @network.each do |tier|
       tier.each do |node|
+        node.sigma = @sigma
         node.mutate
       end
     end
@@ -120,6 +132,8 @@ class NeuralNet
   
   def clone
     clone = NeuralNet.new
+    clone.sigma = self.sigma
+    clone.tau = self.tau
     #iterate in through each tier
     @network.each do |tier|
       nodes = []
@@ -155,6 +169,8 @@ class NeuralNet
   end
   
 protected
+  attr_accessor :sigma, :tau
+  
   def link_tiers
     #first cut lets link every node on a tier to each node on the subsequent tier
     i = 1
@@ -171,4 +187,31 @@ protected
       i +=1
     end
   end
+  
+private
+  #This is a constant related to the size of our network
+  #and the number of connections it contains
+  def recalculate_tau
+    number_of_variables = 0
+    @network.each do |tier|
+      tier.each do |node|
+        number_of_variables += 1 + node.weights.count
+      end
+    end
+    @tau = 1 / (Math.sqrt(2 * Math.sqrt(number_of_variables)))
+  end
+  
+  def gaussian_random
+     u1 = u2 = w = g1 = g2 = 0  # declare
+     begin
+       u1 = 2 * rand - 1
+       u2 = 2 * rand - 1
+       w = u1 * u1 + u2 * u2
+     end while w >= 1
+      
+     w = Math::sqrt( ( -2 * Math::log(w)) / w )
+     g2 = u1 * w;
+     g1 = u2 * w;
+     # g1 is returned  
+  end
 end

data/node.rb

@@ -1,16 +1,18 @@
 class Node
   attr_reader :forward_nodes, :output_value
-  attr_accessor :input_value, :deviation, :weights
-  def initialize
+  attr_accessor :input_value, :sigma, :weights, :bias
+  def initialize sigma
+    @input_value = 0
     @forward_nodes = {}
     @weights =[]
-    @input_value = 0
-    @deviation = rand / 4
+    @bias = 0
+    @sigma = sigma
   end
-  #take our input value, pass it through our sigmoid function (tanh)
-  #and then pass on our output value to each of our forward nodes
+  #take our input value (sum of weighted outputs of backward connected nodes)
+  #, subtract the bias and pass the result through our sigmoid function (tanh)
+  # finally then pass on our output value to each of our forward nodes
   def evaluate
-    @output_value = Math.tanh(@input_value)
+    @output_value = Math.tanh(@input_value - @bias)
     #p "output value #{@output_value}"
     @forward_nodes.each do |node, weight|
       #p "weight #{weight} old input #{node.input_value}"
@@ -22,7 +24,7 @@ class Node
   
   def attach_forward_node node, sequence
     if @weights.count <= sequence
-      @weights << rand
+      @weights << ((rand * 0.4 ) - 0.2) #sampled from a uniform distribution in range ± 0.2
     end
     
     @forward_nodes[node] = @weights[sequence]
@@ -35,7 +37,7 @@ class Node
   def mutate
     new_weights = []
     @weights.each do |weight|
-      new_weights << gaussian_random * @deviation + weight   # new_random_number = gaussian_rand * standard_deviation + average
+      new_weights << weight + @sigma * gaussian_random # new_random_number = average + standard_deviation  * gaussian_rand
     end
     @weights = new_weights
     if @forward_values
@@ -46,17 +48,18 @@ class Node
       end
     end
     
-    #and now mutate the deviation
-    @deviation = (gaussian_random * @deviation)/2 + @deviation
-  end
-  
-  def breed
+    #mutate the bias
+    @bias = @bias + gaussian_random * @sigma
+
+    self
   end
-  
+    
   def clone
-    clone = Node.new
+    clone = Node.new(self.sigma)
+    
+    clone.sigma = self.sigma
+    clone.bias = self.bias
     
-    clone.deviation = self.deviation
     @weights.each do |weight|
       clone.weights << weight
     end
@@ -71,7 +74,7 @@ class Node
   end
   
   def fingerprint
-    fingerprint = "#{@deviation}:"
+    fingerprint = "#{@sigma.to_s}:#{@bias.to_s}:"
     @weights.each do |weight|
       fingerprint += "#{weight.to_s},"
     end
@@ -81,9 +84,10 @@ class Node
   def reload fingerprint
     self.detatch_all_forward_nodes
     @weights = []
-    self.deviation = fingerprint.split(':')[0].to_f
-    if fingerprint.split(":").count == 2
-      fingerprint.split(":")[1].split(',').each do |weight|
+    self.sigma = fingerprint.split(':')[0].to_f
+    self.bias = fingerprint.split(':')[1].to_f
+    if fingerprint.split(":").count == 3
+      fingerprint.split(":")[2].split(',').each do |weight|
         @weights << weight.to_f
       end
     end

metadata

@@ -5,8 +5,8 @@ version: !ruby/object:Gem::Version
   segments: 
   - 0
   - 2
-  - 0
-  version: 0.2.0
+  - 1
+  version: 0.2.1
 platform: ruby
 authors: 
 - David Bochenski
@@ -14,7 +14,7 @@ autorequire:
 bindir: bin
 cert_chain: []
 
-date: 2010-09-04 00:00:00 +01:00
+date: 2010-09-05 00:00:00 +01:00
 default_executable: 
 dependencies: 
 - !ruby/object:Gem::Dependency