Cleanup and reorganization, AI vs AI added

config.py

@@ -5,32 +5,49 @@ class Config(object):
     information parameters. Model objects are passed a Config() object at
     instantiation.
     """
-    models = ['linear', 'dense', 'conv']
-
-    hidden_size = 10
-    # TODO: (feature add) scale the learning rate over time?
-    lr = 0.01
-
+    # Training settings
+    num_iters = 30
     batch_size = 100
     epochs = 50
+    lr = 0.01
 
-    num_layers = 5
-    num_filters = 16
-    kernel_size = 3
-    dropout = 0.1
-
+    # Neural network settings
+    models = ['linear', 'dense', 'conv']
     model = models[2]
     checkpoint_folder = 'checkpoints'
-    arena_games = 40
-    arena_threshold = 0.55
-    # Should be set in a way that encourages exploration in early moves and then 
-    # selects optimal moves later in the game
-    temp_threshold = 6
-    num_iters = 30
+    num_layers = 10
+    hidden_size = 64 # linear and dense only
+    num_filters = 64 # conv only
+    kernel_size = 3 # conv only
+    dropout = 0.1 # conv only
+    regularizer = 0.0001 # resnet only
+    num_residual_blocks = 3 # resnet only
 
-    num_episodes = 100
+    # MCTS settings
+    num_episodes = 25
+    num_sims = 10
     c_puct = 1
-    num_sims = 25
+    # Should be set based on game length to encourage exploration in early moves
+    temp_threshold = 6
+
+    # Unused:
+    #arena_games = 40
+    #arena_threshold = 0.55
+
+# For AI vs AI play
+class Config1(Config):
+    model = Config.models[2]
+    checkpoint_folder = 'checkpoints_old'
+
+    num_layers = 10 # both
+    num_filters = 64 # conv only
+    hidden_size = 64 # linear and dense only
+
+class Config2(Config):
+    model = Config.models[2]
+    checkpoint_folder = 'checkpoints_old'
+
+    num_layers = 15 # both
+    num_filters = 16 # conv only
+    hidden_size = 64 # linear and dense only
 
-    regularizer = 0.0001
-    num_residual_blocks = 3

main.py

@@ -1,28 +1,45 @@
 from retrain import Coach, Individual
-from selfplay import HumanPlay
+from selfplay import Arena, HumanPlay
 from mcts import MCTS
 from zertz.ZertzGame import ZertzGame as Game
 from model import NNetWrapper as NN
-from config import Config
+from config import Config, Config1, Config2
 
 if __name__ == '__main__':
+    # Game settings
     rings = 19
     marbles = {'w': 10, 'g': 10, 'b': 10}
     win_con = [{'w': 2}, {'g': 2}, {'b': 2}, {'w': 1, 'g': 1, 'b': 1}]
     t = 3
 
+    # Setup
     game = Game(rings, marbles, win_con, t)
     config = Config()
     nnet = NN(game, config)
-    trainer = Individual(game, nnet, config)
-
-    # load model weights?
 
-    # learn
+    # Option #1: Learn
+    trainer = Individual(game, nnet, config)
     trainer.learn()
 
-    # play against AI
+    # Option #2: Human vs AI
+    #nnet.load_checkpoint(filename='checkpoint_64_10_29.pth.tar')
     #ai_agent = MCTS(game, nnet, config.c_puct, config.num_sims)
     #hp = HumanPlay(game, ai_agent)
     #hp.play()
 
+    # Option #3: AI vs AI
+    #config1 = Config1()
+    #config2 = Config2()
+    #nnet1 = NN(game, config1)
+    #nnet2 = NN(game, config2)
+
+    #nnet1.load_checkpoint(filename='checkpoint_64_10_29.pth.tar')
+    #nnet2.load_checkpoint(filename='checkpoint_16_15_29.pth.tar')
+
+    #ai_agent1 = MCTS(game, nnet1, config1.c_puct, config1.num_sims)
+    #ai_agent2 = MCTS(game, nnet2, config2.c_puct, config2.num_sims)
+
+    #arena = Arena(game, ai_agent1, ai_agent2)
+    #ai1_win, ai2_win, draw = arena.play_matches(10)
+    #print(ai1_win, ai2_win, draw)
+