update config

src/config.py

@@ -5,26 +5,26 @@
 
 # Model hyperparameters
 nb_filters_3x3 = 64 # number of filters in each 3x3 convolutional layers, 256 in AlphaGo Zero
-value_head_hidden_layer_size = 64 # number of hidden units in the hidden layer of value head
-nb_residual_blocks = 19 # number of residual blocks in the neural network architecture
+value_head_hidden_layer_size = 128 # number of hidden units in the hidden layer of value head
+nb_residual_blocks = 10 # number of residual blocks in the neural network architecture
 
 # MCTS parameters
 max_iter_move_to_leaf = 300 # one argument for mcts.move_to_leaf()
-cpuct = 0.02 # exploration / exploitation
+cpuct = 0.1 # exploration / exploitation
 
 # Training parameters
-lr = 0.001
-random_seed = 42
-batch_size = 32
+lr = 0.01
+random_seed = 0
+batch_size = 64
 nb_iter_training = 1000
 
 # Pipeline parameters
-nb_epochs_per_iteration = 2 # Number of epochs of training of ConvNet in every iteration of pipeline
-dataset_max_size = 5000 # Maximum number of kept self-play records, 500000 in AlphaGo Zero
-resignation_threshold = - 0.2 # Resignation if the estimated value is lower than this threshold
-nb_simulations = 2000 # Number of simulations for MCTS
-nb_self_play_in_each_iteration = 3 # Number of self-play games in each iteration of training, 25000 in AlphaGo Zero
-freq_iter_checkpoint = 10 # Frequency of saving checkpoint by iteration, 1000 in AlphaGo Zero's paper
+nb_epochs_per_iteration = 50 # Number of epochs of training of ConvNet in every iteration of pipeline
+dataset_max_size = 20000 # Maximum number of kept self-play records, 500000 in AlphaGo Zero
+resignation_threshold = - 0.75 # Resignation if the estimated value is lower than this threshold
+nb_simulations = 100 # Number of simulations for MCTS
+nb_self_play_in_each_iteration = 10 # Number of self-play games in each iteration of training, 25000 in AlphaGo Zero
+freq_iter_checkpoint = 100 # Frequency of saving checkpoint by iteration, 1000 in AlphaGo Zero's paper
 freq_update_best_mcts = 1 # Frequency of updating the best player so far, 1000 in AlphaGo Zero
 train_net_log_interval = 10 # Interval (batch_idx) of print