class ReinforcementLearner:
__metaclass__ = abc.ABCMeta
lock = threading.Lock()
def __init__(self,
rl_method='rl',
stock_code=None,
chart_data=None,
training_data=None,
min_trading_unit=1,
max_trading_unit=2,
delayed_reward_threshold=.05,
net='dnn',
num_steps=1,
lr=0.001,
value_network=None,
policy_network=None,
output_path='',
reuse_models=True):
# 인자 확인
assert min_trading_unit > 0
assert max_trading_unit > 0
assert max_trading_unit >= min_trading_unit
assert num_steps > 0
assert lr > 0
# 강화학습 기법 설정
self.rl_method = rl_method
# 환경 설정
self.stock_code = stock_code
self.chart_data = chart_data
self.environment = Environment(chart_data)
# 에이전트 설정
self.agent = Agent(self.environment,
min_trading_unit=min_trading_unit,
max_trading_unit=max_trading_unit,
delayed_reward_threshold=delayed_reward_threshold)
# 학습 데이터
self.training_data = training_data
self.sample = None
self.training_data_idx = -1
# 벡터 크기 = 학습 데이터 벡터 크기 + 에이전트 상태 크기
self.num_features = self.agent.STATE_DIM
if self.training_data is not None:
self.num_features += self.training_data.shape[1]
# 신경망 설정
self.net = net
self.num_steps = num_steps
self.lr = lr
self.value_network = value_network
self.policy_network = policy_network
self.reuse_models = reuse_models
self.critic = value_network
self.actor = policy_network
self.tau = 0.001
# 가시화 모듈
self.visualizer = Visualizer()
# 메모리
self.memory_sample = []
self.memory_action = []
self.memory_reward = []
self.memory_value = []
self.memory_policy = []
self.memory_target_policy = []
self.memory_target_value = []
self.memory_target_action = []
self.memory_pv = []
self.memory_num_stocks = []
self.memory_exp_idx = []
self.memory_learning_idx = []
# 에포크 관련 정보
self.loss = 0.
self.itr_cnt = 0
self.exploration_cnt = 0
self.batch_size = 0
self.learning_cnt = 0
# 로그 등 출력 경로
self.output_path = output_path
def init_policy_network(self,
shared_network=None,
activation='sigmoid',
loss='binary_crossentropy'):
if self.rl_method == 'ddpg':
print("actor")
self.actor = ActorNetwork(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
num_steps=self.num_steps,
activation=activation,
loss=loss,
lr=self.lr)
print(self.actor)
elif self.net == 'dnn':
self.policy_network = DNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'lstm':
self.policy_network = LSTMNetwork(
input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'cnn':
self.policy_network = CNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'cnn':
self.policy_network = CNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
if self.reuse_models and \
os.path.exists(self.policy_network_path):
self.policy_network.load_model(model_path=self.policy_network_path)
def init_value_network(self,
shared_network=None,
activation='linear',
loss='mse'):
if self.rl_method == 'ddpg':
self.critic = CriticNetwork(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
num_steps=self.num_steps,
activation=activation,
loss=loss,
lr=self.lr)
elif self.net == 'dnn':
self.value_network = DNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'lstm':
self.value_network = LSTMNetwork(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'cnn':
self.value_network = CNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
elif self.net == 'cnn':
self.value_network = CNN(input_dim=self.num_features,
output_dim=self.agent.NUM_ACTIONS,
lr=self.lr,
num_steps=self.num_steps,
shared_network=shared_network,
activation=activation,
loss=loss)
if self.reuse_models and \
os.path.exists(self.value_network_path):
self.value_network.load_model(model_path=self.value_network_path)
def reset(self):
self.sample = None
self.training_data_idx = -1
# 환경 초기화
self.environment.reset()
# 에이전트 초기화
self.agent.reset()
# 가시화 초기화
self.visualizer.clear([0, len(self.chart_data)])
# 메모리 초기화
self.memory_sample = []
self.memory_action = []
self.memory_target_policy = []
self.memory_target_value = []
self.memory_target_action = []
self.memory_reward = []
self.memory_value = []
self.memory_policy = []
self.memory_pv = []
self.memory_num_stocks = []
self.memory_exp_idx = []
self.memory_learning_idx = []
# 에포크 관련 정보 초기화
self.loss = 0.
self.itr_cnt = 0
self.exploration_cnt = 0
self.batch_size = 0
self.learning_cnt = 0
def build_sample(self):
self.environment.observe()
if len(self.training_data) > self.training_data_idx + 1:
self.training_data_idx += 1
self.sample = self.training_data.iloc[
self.training_data_idx].tolist()
self.sample.extend(self.agent.get_states())
return self.sample
return None
@abc.abstractmethod
def get_batch(self, batch_size, delayed_reward, discount_factor):
pass
@abc.abstractmethod
def train(self, batch_size, delayed_reward, discount_factor):
pass
def update_networks(self, batch_size, delayed_reward, discount_factor):
# 배치 학습 데이터 생성
x, y_value, y_policy = self.get_batch(batch_size, delayed_reward,
discount_factor)
if len(x) > 0:
loss = 0
if y_value is not None:
# 가치 신경망 갱신
loss += self.critic.train_on_batch(x, y_value)
self.critic.transfer_weights()
if y_policy is not None:
# 정책 신경망 갱신
loss += self.actor.train_on_batch(x, y_policy)
self.actor.transfer_weights()
return loss
return None
def fit(self, delayed_reward, discount_factor, full=False):
batch_size = len(self.memory_reward) if full \
else self.batch_size
# 배치 학습 데이터 생성 및 신경망 갱신
if batch_size > 0:
_loss = self.update_networks(batch_size, delayed_reward,
discount_factor)
if _loss is not None:
self.loss += abs(_loss)
self.learning_cnt += 1
self.memory_learning_idx.append(self.training_data_idx)
self.batch_size = 0
def visualize(self, epoch_str, num_epoches, epsilon):
self.memory_action = [Agent.ACTION_HOLD] \
* (self.num_steps - 1) + self.memory_action
self.memory_num_stocks = [0] * (self.num_steps - 1) \
+ self.memory_num_stocks
if self.value_network is not None:
self.memory_value = [np.array([np.nan] \
* len(Agent.ACTIONS))] * (self.num_steps - 1) \
+ self.memory_value
if self.policy_network is not None:
self.memory_policy = [np.array([np.nan] \
* len(Agent.ACTIONS))] * (self.num_steps - 1) \
+ self.memory_policy
self.memory_pv = [self.agent.initial_balance] \
* (self.num_steps - 1) + self.memory_pv
self.visualizer.plot(
epoch_str=epoch_str,
num_epoches=num_epoches,
epsilon=epsilon,
action_list=Agent.ACTIONS,
actions=self.memory_action,
num_stocks=self.memory_num_stocks,
outvals_value=self.memory_value,
outvals_policy=self.memory_policy,
exps=self.memory_exp_idx,
learning_idxes=self.memory_learning_idx,
initial_balance=self.agent.initial_balance,
pvs=self.memory_pv,
)
self.visualizer.save(
os.path.join(self.epoch_summary_dir,
'epoch_summary_{}.png'.format(epoch_str)))
def run(self,
num_epoches=100,
balance=10000000,
discount_factor=0.9,
start_epsilon=0.5,
learning=True):
info = "[{code}] RL:{rl} Net:{net} LR:{lr} " \
"DF:{discount_factor} TU:[{min_trading_unit}," \
"{max_trading_unit}] DRT:{delayed_reward_threshold}".format(
code=self.stock_code, rl=self.rl_method, net=self.net,
lr=self.lr, discount_factor=discount_factor,
min_trading_unit=self.agent.min_trading_unit,
max_trading_unit=self.agent.max_trading_unit,
delayed_reward_threshold=self.agent.delayed_reward_threshold
)
with self.lock:
logging.info(info)
# 시작 시간
time_start = time.time()
# 가시화 준비
# 차트 데이터는 변하지 않으므로 미리 가시화
self.visualizer.prepare(self.environment.chart_data, info)
# 가시화 결과 저장할 폴더 준비
self.epoch_summary_dir = os.path.join(
self.output_path, 'epoch_summary_{}'.format(self.stock_code))
if not os.path.isdir(self.epoch_summary_dir):
os.makedirs(self.epoch_summary_dir)
else:
for f in os.listdir(self.epoch_summary_dir):
os.remove(os.path.join(self.epoch_summary_dir, f))
# 에이전트 초기 자본금 설정
self.agent.set_balance(balance)
# 학습에 대한 정보
max_portfolio_value = 0
epoch_win_cnt = 0
# 학습 반복
for epoch in range(num_epoches):
time_start_epoch = time.time()
# step 샘플을 만들기 위한 큐
q_sample = collections.deque(maxlen=self.num_steps)
# 환경, 에이전트, 신경망, 가시화, 메모리 초기화
self.reset()
# 학습을 진행할 수록 탐험 비율 감소
if learning:
epsilon = start_epsilon \
* (1. - float(epoch) / (num_epoches - 1))
self.agent.reset_exploration()
else:
epsilon = start_epsilon
while True:
# 샘플 생성
next_sample = self.build_sample()
if next_sample is None:
break
# num_steps만큼 샘플 저장
q_sample.append(next_sample)
if len(q_sample) < self.num_steps:
continue
# 가치, 정책 신경망 예측
pred_value = None
pred_policy = None
pred_target_policy = None
pred_target_value = None
if self.critic is not None:
pred_value = self.critic.predict(list(q_sample))
pred_target_value = self.critic.target_predict(
list(q_sample))
if self.actor is not None:
pred_policy = self.actor.predict(list(q_sample))
pred_target_policy = self.actor.target_predict(
list(q_sample))
# 신경망 또는 탐험에 의한 행동 결정
action, confidence, exploration = \
self.agent.decide_action(pred_value, pred_policy, epsilon)
# target 값을 이용한 행동 결정
target_action, target_confidence, target_exploration = \
self.agent.decide_action(pred_target_policy, pred_target_value, epsilon)
#결정한 행동을 수행하고 즉시 보상과 지연 보상 획득
immediate_reward, delayed_reward = \
self.agent.act(action, confidence)
# 행동 및 행동에 대한 결과를 기억
self.memory_sample.append(list(q_sample))
self.memory_action.append(action)
self.memory_reward.append(immediate_reward)
self.memory_target_action.append(target_action)
self.memory_target_policy.append(pred_target_policy)
self.memory_target_value.append(pred_target_value)
if self.value_network is not None:
self.memory_value.append(pred_value)
if self.policy_network is not None:
self.memory_policy.append(pred_policy)
self.memory_pv.append(self.agent.portfolio_value)
self.memory_num_stocks.append(self.agent.num_stocks)
if exploration:
self.memory_exp_idx.append(self.training_data_idx)
# 반복에 대한 정보 갱신
self.batch_size += 1
self.itr_cnt += 1
self.exploration_cnt += 1 if exploration else 0
# 지연 보상 발생된 경우 미니 배치 학습
if learning and (delayed_reward != 0):
self.fit(delayed_reward, discount_factor)
# 에포크 종료 후 학습
if learning:
self.fit(self.agent.profitloss, discount_factor, full=True)
# 에포크 관련 정보 로그 기록
num_epoches_digit = len(str(num_epoches))
epoch_str = str(epoch + 1).rjust(num_epoches_digit, '0')
time_end_epoch = time.time()
elapsed_time_epoch = time_end_epoch - time_start_epoch
if self.learning_cnt > 0:
logging.info("[{}][Epoch {}/{}] Epsilon:{:.4f} "
"#Expl.:{}/{} #Buy:{} #Sell:{} #Hold:{} "
"#Stocks:{} PV:{:,.0f} "
"LC:{} Loss:{:.6f} ET:{:.4f}".format(
self.stock_code, epoch_str, num_epoches,
epsilon, self.exploration_cnt, self.itr_cnt,
self.agent.num_buy, self.agent.num_sell,
self.agent.num_hold, self.agent.num_stocks,
self.agent.portfolio_value, self.learning_cnt,
self.loss, elapsed_time_epoch))
# 에포크 관련 정보 가시화
self.visualize(epoch_str, num_epoches, epsilon)
# 학습 관련 정보 갱신
max_portfolio_value = max(max_portfolio_value,
self.agent.portfolio_value)
if self.agent.portfolio_value > self.agent.initial_balance:
epoch_win_cnt += 1
# 종료 시간
time_end = time.time()
elapsed_time = time_end - time_start
# 학습 관련 정보 로그 기록
with self.lock:
logging.info("[{code}] Elapsed Time:{elapsed_time:.4f} "
"Max PV:{max_pv:,.0f} #Win:{cnt_win}".format(
code=self.stock_code,
elapsed_time=elapsed_time,
max_pv=max_portfolio_value,
cnt_win=epoch_win_cnt))
def save_models(self):
if self.value_network is not None and \
self.value_network_path is not None:
self.value_network.save_model(self.value_network_path)
if self.policy_network is not None and \
self.policy_network_path is not None:
self.policy_network.save_model(self.policy_network_path)
class DDPG(object):
"""Implementation of the deep deterministic policy gradient algorithm"""
def __init__(self,
docker_client,
name='worker',
port=3101,
model_path='../models/ddpg',
log_path='../logs/ddpg'):
self.state_size = 29
self.action_size = 3
self.docker_client = docker_client
self.buffer_size = 100000
self.batch_size = 32
self.gamma = 0.99 # disocunt factor
self.tau = 0.001 # Target Network HyperParameters
self.lra = 0.0001 # Learning rate for Actor
self.lrc = 0.001 # Lerning rate for Critic
seed(6486)
self.explore = 100000.
self.episode_count = 2000
self.max_steps = 10000
self.epsilon = 1
self.model_path = model_path
self.port = port
self.name = name
if not os.path.exists(self.model_path):
os.makedirs(self.model_path)
self.config = tf.ConfigProto()
self.config.gpu_options.allow_growth = True
tf.reset_default_graph()
self.summary_writer = tf.summary.FileWriter(log_path)
self.actor = ActorNetwork(self.state_size, self.action_size,
tf.train.AdamOptimizer(self.lra), self.tau)
self.critic = CriticNetwork(self.state_size, self.action_size,
tf.train.AdamOptimizer(self.lrc), self.tau)
self.buff = ReplayBuffer(self.buffer_size)
self.saver = tf.train.Saver()
self._create_summary()
self.summary_histogram = tf.summary.merge_all()
def _create_summary(self):
with tf.name_scope('summary'):
self.loss_summary_op = tf.summary.scalar('loss',
self.critic.loss,
collections=['loss'])
self.reward_ph = tf.placeholder(shape=[
None,
],
name='reward',
dtype=tf.float32)
self.target_q_values_ph = tf.placeholder(
shape=[None, self.action_size],
name='target_q_values',
dtype=tf.float32)
self.y_t_ph = tf.placeholder(shape=[None, self.action_size],
name='target_y_t',
dtype=tf.float32)
tf.summary.scalar('reward',
tf.reduce_mean(self.reward_ph),
collections=['reward'])
tf.summary.scalar('target_q_values',
tf.reduce_mean(self.target_q_values_ph),
collections=['reward'])
tf.summary.scalar('y_t',
tf.reduce_mean(self.y_t_ph),
collections=['reward'])
self.reward_summary_op = tf.summary.merge_all('reward')
@staticmethod
def addOUNoise(a, epsilon):
"""Adds noise from an Ornstein Uhlenbeck process to the actions"""
def ou_func(x, mu, theta, sigma):
return theta * (mu - x) + sigma * randn(1)
a_new = np.zeros(np.shape(a))
noise = np.zeros(np.shape(a))
noise[0] = (max(epsilon, 0) * ou_func(a[0], 0.0, 0.60, 0.30))
noise[1] = (max(epsilon, 0) * ou_func(a[1], 0.5, 1.00, 0.10))
noise[2] = (max(epsilon, 0) * ou_func(a[2], -0.1, 1.00, 0.10))
a_new[0] = a[0] + noise[0]
a_new[1] = a[1] + noise[1]
a_new[2] = a[2] + noise[2]
return a_new
def train(self, track_name='', check_stuck=True):
all_steps = 0
if track_name == '':
env = TorcsDockerEnv(self.docker_client,
self.name,
self.port,
training=True)
else:
env = TorcsDockerEnv(self.docker_client,
self.name,
self.port,
track_name=track_name)
with tf.Session(config=self.config) as sess:
sess.run(tf.global_variables_initializer())
ckpt = tf.train.latest_checkpoint(self.model_path)
if ckpt:
print('load model weights from {}'.format(ckpt))
self.saver.restore(sess, ckpt)
for i in range(self.episode_count):
# collect the recent rewards
recent_rewards = np.ones(1000) * 1e9
print("Episode : " + str(i) + " Replay Buffer " +
str(self.buff.count()))
if np.mod(i, 3) == 0:
observation = env.reset(relaunch=True)
else:
observation = env.reset()
state_t = obs_to_state(observation)
total_reward = 0
for j in range(self.max_steps):
loss = 0
# reduce the effect of the OU process with progess in the
# algorithm
self.epsilon -= 1.0 / self.explore
action_t = self.actor.predict(
sess, state_t.reshape(1, state_t.shape[0]))
observation, reward_t, done, _ = env.step(
DDPG.addOUNoise(action_t[0], self.epsilon))
state_t1 = obs_to_state(observation)
# check if we need to terminate, bc the agent is stuck
recent_rewards[j % 1000] = reward_t
if (check_stuck and np.median(recent_rewards) < 1.0
and i / self.episode_count < 0.5):
break
self.buff.add(state_t, action_t[0], reward_t, state_t1,
done)
batch = self.buff.getBatch(self.batch_size)
states = np.asarray([e[0] for e in batch])
actions = np.asarray([e[1] for e in batch])
rewards = np.asarray([e[2] for e in batch])
new_states = np.asarray([e[3] for e in batch])
dones = np.asarray([e[4] for e in batch])
y_t = np.asarray([e[1] for e in batch])
target_q_values = self.critic.target_predict(
sess, new_states,
self.actor.target_predict(sess, new_states))
for k in range(len(batch)):
if dones[k]:
y_t[k] = rewards[k]
else:
y_t[k] = (rewards[k] +
self.gamma * target_q_values[k])
loss += self.critic.train(sess, y_t, states, actions)
actions_for_grad = self.actor.predict(sess, states)
grads = self.critic.gradients(sess, states,
actions_for_grad)
self.actor.train(sess, states, grads)
self.actor.target_train(sess)
self.critic.target_train(sess)
all_steps += 1
if j % 50:
loss_summary, reward_summary, histogram = sess.run(
[
self.loss_summary_op, self.reward_summary_op,
self.summary_histogram
],
feed_dict={
self.critic.expected_critic: y_t,
self.critic.state: states,
self.actor.state: states,
self.actor.target_state: states,
self.critic.action: actions,
self.reward_ph: rewards,
self.target_q_values_ph: target_q_values,
self.y_t_ph: y_t
})
self.summary_writer.add_summary(
loss_summary, all_steps)
self.summary_writer.add_summary(
reward_summary, all_steps)
self.summary_writer.add_summary(histogram, all_steps)
self.summary_writer.flush()
total_reward += reward_t
state_t = state_t1
print("Episode", i, "Step", all_steps, "Action", action_t,
"Reward", reward_t, "Loss", loss)
if done:
break
print("TOTAL REWARD @ " + str(i) + "-th Episode : Reward " +
str(total_reward))
print("Total Step: " + str(all_steps))
print("")
if np.mod(i, 50) == 0:
self.saver.save(
sess, self.model_path + '/model-{:d}.cptk'.format(i))
env.end()