create_q_model = __import__('train').create_q_model
Atari_Processor = __import__('train').Atari_Processor
if __name__ == '__main__':
"""PLay an episode of the game"""
windowLen = 4
# get the environment and the nuber of actions
env = gym.make("Breakout-v0")
env.reset()
num_actions = env.action_space.n
model = create_q_model(num_actions, windowLen)
memory = SequentialMemory(limit=1000000, window_length=windowLen)
processor = AtariProcessor()
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(), attr='eps', value_max=1,
value_min=.1, value_test=.05,
nb_steps=1000000)
# training the model
dqn = DQNAgent(model=model, nb_actions=num_actions,
policy=policy, memory=memory,
processor=processor)
dqn.cmopile(Adam(lr=.00025), metrics=['mae'])
# loading weights after training is done
dqn.load_weights('policy.h5')
# Evaluating algorithm
dqn.test(env, nb_episodes=10, visualize=True)
