var for var in tf.global_variables() if var.name[:9] != "embedding"
])
restore_path = tf.train.latest_checkpoint(logdir)
with tf.Session() as sess:
if restore_path is not None:
logger.info(
"Restoring variables from checkpoint: {}".format(restore_path))
restoring_saver.restore(sess, restore_path)
else:
logger.info("Initializing brand new network parameters.")
sess.run(tf.global_variables_initializer())
global_step = sess.run(ml.global_step)
logger.info("Gathering initial gameplay data!")
if config['training_agent'] == "random_rollout":
from agents import RandomRolloutAgent
agent = RandomRolloutAgent(ml.envmodel)
policy = agent.policy
else:
policy = None
ml.gather_gameplay_data(config['n_initial_games'], policy=None)
logger.info("Beginning training.")
logger.info("To visualize, call:\ntensorboard --logdir={}".format(logdir))
from utils import dataset
while (not config['maxsteps']) or global_step < config['maxsteps']:
transition_data = ml.create_transition_sequence_dataset(n=10000)
if config['use_goal_boosting']:
gb_data = ml.create_goals_dataset(n=1000)
gb_data_batches = dataset.iterbatches(
gb_data,
batch_size=config['batchsize'],
shuffle=True,
from agents import RandomRolloutAgent
from environments.connect4 import Connect4GameState
from runners import run_to_the_end, run_for_n_games_and_print_stats
if __name__ == "__main__":
gs = Connect4GameState()
agent0 = RandomRolloutAgent(10, False)
agent1 = RandomRolloutAgent(10, False)
run_for_n_games_and_print_stats([agent0, agent1], gs, 10)
from agents import CommandLineAgent, DeepQLearningAgent, RandomRolloutAgent
from environments.connect4 import Connect4GameState
from runners import run_to_the_end, run_for_n_games_and_print_stats, run_step
if __name__ == "__main__":
gs = Connect4GameState()
agent0 = DeepQLearningAgent(action_space_size=gs.get_action_space_size(),
neurons_per_hidden_layer=128,
hidden_layers=5)
agent1 = RandomRolloutAgent(100, False)
agent0.alpha = 0.1
agent0.epsilon = 0.3
for i in range(100):
run_for_n_games_and_print_stats([agent0, agent1], gs, 100)
agent0.epsilon = -1.0
run_for_n_games_and_print_stats([agent0, agent1], gs, 100)
gs_clone = gs.clone()
while not gs_clone.is_game_over():
run_step([agent0, CommandLineAgent()], gs_clone)
print(gs_clone)
from agents import RandomRolloutAgent
from environments import GridWorldGameState
from runners import run_to_the_end
if __name__ == "__main__":
gs = GridWorldGameState()
agent = RandomRolloutAgent(100000, True)
print(gs)
run_to_the_end([agent], gs)
print(gs)