def main():
nb_actions = 3
obs_size = 9
window_size = 10
batch_size = 2048
stock = "BAC"
episode = 35
total_spent = 0
total_sold = 0
agent = Agent(window_size=window_size,
action_size=nb_actions,
batch_size=batch_size,
gamma=0.95,
epsilon=1.0,
epsilon_decay=0.99,
epsilon_min=0.001,
learning_rate=0.001,
is_eval=True,
stock_name=stock,
episode=episode)
env = MarketEnv(stock,
window_size=window_size,
state_size=obs_size,
account_balance=1000000,
is_eval=True,
shares_to_buy=100,
max_positions=1000,
train_test_split=.8)
state = env.reset()
for time in range(env.l):
action = agent.act(state)[0]
if action < 0:
choice = 2
elif action > 0 and action[0] < 1:
choice = 0
elif action > 1:
choice = 1
next_state, action, reward, done = env.step(action, time)
agent.remember(state, action, reward, next_state, done)
state = next_state
prices = [line[3] for line in env.prices]
dates = [i for i in range(len(env.prices))]
plt.plot(dates, prices)
for line in env.buy:
plt.plot(line[0], line[1], 'ro', color="g", markersize=2)
total_spent += line[1]
for line in env.sell:
plt.plot(line[0], line[1], "ro", color="r", markersize=2)
total_sold += line[1]
percentage_gain = ((env.account_balance - env.starting_balance) /
env.starting_balance) * 100
print("Profitable Trades: " + str(env.profitable_trades))
print("Unprofitable Trades: " + str(env.unprofitable_trades))
print("Percentage Gain: " + str(percentage_gain))
print("Amount Spent: " + str(total_spent))
print("Amount Sold: " + str(total_sold))
plt.show()
plt.savefig("models/{}/{}-{}/{}".format(stock, stock, str(episode), stock))
for e in range(epoch):
#loss = 0.
game_over = False
input_t = env.reset()
print("run status {}", format(e))
while not game_over:
input_prev = input_t
#print(input_prev)
isRandom = False
if np.random.rand() <= epsilon:
action = np.random.randint(0, env.action_space.n, size=1)[0]
isRandom = True
else:
q = model.predict(np.array([input_prev]))
action = np.argmax(q[0])
input_t, reward, game_over, info = env.step(action)
if game_over == True:
print("total reward : ", env.cur_reward)
exp_replay.remember([input_prev, action, reward, input_t], game_over)
batch = exp_replay.get_batch(model, batch_size=batch_size)
# loss = model.train_on_batch(batch[0], batch[1])
# model_json = model.to_json()
# with open("model.json", "w") as json_file:
# json_file.write(model_json)
# model.save_weights("model.h5")
# model.save_weights("model_bk.h5")
def main():
window_size = 10
batch_size = 2048
episodes = 10000
max_episode_len = 39000 * 3 # One Year of trading in minutes
stock = "BAC"
args = {
'tau': .001,
'gamma': .99,
'lr_actor': .0001,
'lr_critic': .001,
'batch_size': max_episode_len
}
env = MarketEnv(stock,
buy_position=3,
window_size=window_size,
account_balance=1000000,
shares_to_buy=100,
train_test_split=.8,
max_episode_len=max_episode_len)
agent = Agent(args,
state_size=env.state_size,
window_size=env.window_size,
action_size=env.action_size,
action_bound=env.action_bound[1],
is_eval=False,
stock_name=stock)
episode_ave_max_q = 0
ep_reward = 0
for i in range(episodes):
state = env.reset()
for time in range(env.l):
action = agent.act(state)[0]
if action < 0:
choice = 2
elif action > 0 and action[0] < 1:
choice = 0
elif action > 1:
choice = 1
next_state, reward, done = env.step(choice, time)
agent.remember(state, action, reward, next_state, done)
state = next_state
# if agent.replay_buffer.size() == batch_size:
# print("Replaying")
# episode_ave_max_q += agent.replay(time, i, episode_ave_max_q)
ep_reward += reward
if done or time == env.l:
episode_ave_max_q += agent.replay(time, i, episode_ave_max_q)
break
model_name = "{}-{}".format(stock, str(i))
path = "models/{}/{}/".format(stock, model_name)
if i % 5 == 0:
if not os.path.exists(path):
os.makedirs(path)
with open(os.path.join(path, 'LTYP.mif'), 'w'):
pass
agent.saver.save(agent.sess, path + model_name, global_step=i)
summary_str = agent.sess.run(agent.summary_ops,
feed_dict={
agent.summary_vars[0]:
ep_reward,
agent.summary_vars[1]:
episode_ave_max_q
})
agent.writer.add_summary(summary_str, i)
agent.writer.flush()
episode_ave_max_q = 0
ep_reward = 0
print('| Reward: {:d} | Episode: {:d} | Qmax: {:.4f}'.format(
int(ep_reward), i, (episode_ave_max_q)))