memory = SequentialMemory(limit=50000, window_length=window)
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(),
attr='eps',
value_max=1.,
value_min=.1,
value_test=.05,
nb_steps=5000)
#policy = EpsGreedyQPolicy(10)
#policy = BoltzmannQPolicy()
dqn = DQNAgent(model=model,
nb_actions=num_actions,
memory=memory,
nb_steps_warmup=window * 3,
policy=policy)
dqn.compile(Adam(lr=1e-3), metrics=['mae'])
dqn.enable_dueling_network = True
if model_exist:
dqn.load_weights(model_path)
dqn.policy = LinearAnnealedPolicy(EpsGreedyQPolicy(),
attr='eps',
value_max=.5,
value_min=.1,
value_test=.05,
nb_steps=5000)
env.set_data_interval(train_start, train_end)
train_history = dqn.fit(env,
nb_steps=5000,
visualize=False,
verbose=2,
action_repetition=5)
