def run_maze():
step = 0
for episode in range(5):
# initial observation
observation = env.reset()
while True:
# fresh env
#env.render()
# RL choose action based on observation
action = DQN.choose_action(observation)
# RL take action and get next observation and reward
observation_, reward, done = env.step(action)
DQN.store_transition(observation, action, reward, observation_)
if (step > 200) and (step % 5 == 0):
DQN.learn()
# swap observation
observation = observation_
# break while loop when end of this episode
if done:
print('game over')
print(env.balance)
break
step += 1
# action_low = 1
# observation_, reward, done, info = env.step(action_high=action_high, action_low=action_low)
# flag = 1
# else:
# action_high = 0
# action_low = 1
# observation_, reward, done, info = env.step(action_high=action_high, action_low=action_low)
# flag = 0
observation_ = np.array(observation_)
print("action_high: " + str(action_high) + " action_low: " +
str(action_low))
print("reward: " + str(reward))
print("info: " + str(info))
print("-------------------------")
if (step + 1) % 100 == 0:
elapsed = (time.clock() - start)
print("Step: " + str(step + 1) + " Time Used: " + str(elapsed))
observation = observation_
step += 1
if __name__ == "__main__":
LC_env = Env(ego_start_time=100)
# Just show
dqn = DQN(n_features=6,
e_greedy_start=1,
e_greedy_increment=0.01,
is_save=True,
is_restore=True)
run_task(env=LC_env, no_gui=False, max_episode=1, net=dqn)
import gym
from RL_brain import DQN
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
env = gym.make('MountainCar-v0')
env = env.unwrapped
env.seed(21)
MEMORY_SIZE = 10000
sess = tf.Session()
with tf.variable_scope('natural_DQN'):
RL_natural = DQN(
n_actions=3, n_features=2, memory_size=MEMORY_SIZE,
e_greedy_increment=0.00005, sess=sess, prioritized=False,
)
with tf.variable_scope('DQN_with_prioritized_replay'):
RL_prio = DQN(
n_actions=3, n_features=2, memory_size=MEMORY_SIZE,
e_greedy_increment=0.00005, sess=sess, prioritized=True, output_graph=True,
)
sess.run(tf.global_variables_initializer())
def train(RL):
total_steps = 0
steps = []
episodes = []
for i_episode in range(20):
# RL choose action based on observation
action = RL.choose_action(observation)
# RL take action and get next observation and reward
observation_, reward, done = env.step(action)
RL.store_transition(observation, action, reward, observation_)
if (step > 200) and (step % 5 == 0):
RL.learn()
# swap observation
observation = observation_
# break while loop when end of this episode
if done:
break
step += 1
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
# maze game
env = Maze()
RL = DQN()
env.after(100, run_maze)
env.mainloop()
import gym
from RL_brain import DQN
env = gym.make('MountainCar-v0')
print(env.action_space)
print(env.observation_space)
print(env.observation_space.high)
print(env.observation_space.low)
RL = DQN(n_states=2,
n_actions=3,
learning_rate=0.001,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=300,
memory_size=3000,
e_greedy_increment=0.002)
total_steps = 0
for i_episode in range(100):
observation = env.reset()
ep_r = 0
while True:
env.render()
action = RL.choose_action(observation)
observation_, reward, done, info = env.step(action)
s_, reward, done = env.step(action)
# RL learn from this transition
RL.learn(s, action, reward, s_)
if len(RL.memory) >= 200:
training = np.array(RL.memory)
RL.q_model.fit(np.array(training[:, 0]),
np.array([i for i in training[:, 1]]),
epochs=5)
RL.memory = []
# swap observation
s = s_
# break while loop when end of this episode
if done:
#RL.epsilon += 0.001
break
G = GrowUp()
print("test")
for i in range(env.fin_step):
q_table = RL.q_model.predict([i])
G.step(np.argmax(q_table))
print(G.score)
if __name__ == "__main__":
env = GrowUp()
RL = DQN(actions=list(range(env.n_actions)))
update()
observation = observation_
# break while loop when end of this episode
if done == 1:
print('win!')
break
elif done == -1:
print('lose!')
break
step += 1
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
env = Maze()
RL = DQN(env.n_features,
env.n_actions,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=200,
memory_size=2000)
env.after(100, update)
env.mainloop()
RL.plot_cost()
import gym
from RL_brain import DQN
env = gym.make('CartPole-v0')
print(env.action_space)
print(env.observation_space)
print(env.observation_space.high)
print(env.observation_space.low)
RL = DQN(
env.observation_space.shape[0],
env.action_space.n,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=100,
memory_size=2000,
e_greedy_increment=0.001)
total_steps = 0
for i_episode in range(100):
observation = env.reset()
ep_r = 0
while True:
env.render()
action = RL.choose_action(observation)
observation = observation_
# break while loop when end of this episode
if done:
print('game over')
print(env.balance)
break
step += 1
# end of game
#env.destroy()
if __name__ == "__main__":
# maze game
env = FX()
DQN = DQN(
env.n_actions,
env.n_features,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=200,
memory_size=2000,
# output_graph=True
)
# env.after(100, run_maze)
run_maze()
# env.mainloop()
DQN.plot_cost()
def run_model(_train_episode=100,
_learn_threshold=200,
_learn_interval=5,
_base_currency='USD',
_trade_log_mode='NONE',
_trade_log_raw=False,
_trade_log_to_file=False,
_show_checkout=True,
_show_step=True):
step = 0
for episode in range(_train_episode):
observation, TI_initial, initial_time = env.reset()
TI_initial_balance = copy.deepcopy(TI_initial)
train_name = TI_initial.account_name
while True:
action = DQN.choose_action(observation)
observation_, reward, done, TI_end, end_time = env.step(
action, print_step=_show_step)
DQN.store_transition(observation, action, reward, observation_)
if (step > _learn_threshold) and (step % _learn_interval == 0):
DQN.learn()
observation = observation_
if done:
print('$' * 20 + ' GAME OVER ' + '$' * 20)
TI_initial_balance.account_name = 'Initial_Checkout_Review' + ' (episode: ' + str(
episode + 1) + ')'
TI_initial_balance.checkout_all_in(initial_time,
_base_currency)
TI_end_balance = copy.deepcopy(TI_end)
TI_end_balance.account_name = 'End_Checkout_Review' + ' (episode: ' + str(
episode + 1) + ')'
TI_end_balance.checkout_all_in(end_time, _base_currency)
print('MI: Initial Time: {}; End Time: {}'.format(
initial_time, end_time))
if _show_checkout == True:
TI_initial_balance.account_review()
TI_end_balance.account_review()
if _trade_log_mode == False:
pass
elif _trade_log_mode == 'ALL':
TI_end_balance.trade_log_review(raw_flag=_trade_log_raw)
elif _trade_log_mode == 'TWOENDS':
TI_end.trade_log_review(tar_action_id=0,
raw_flag=_trade_log_raw)
TI_end.trade_log_review(tar_action_id='LAST',
raw_flag=_trade_log_raw)
else:
print(
'Invalid _trade_log_mode input ({}). Must be \'ALL\', \'TWOENDS\', or False'
.format(_trade_log_mode))
return -1
if _trade_log_to_file:
trade_log_base_dir = './logs/trade_logs/'
if not os.path.exists(trade_log_base_dir):
os.makedirs(trade_log_base_dir)
trade_log_file_name = trade_log_base_dir + str(train_name)
log_file_readable = open(trade_log_file_name + '.txt',
'w+')
with contextlib.redirect_stdout(log_file_readable):
TI_end.trade_log_review()
print(
"### READABLE trade_log of {} successfully exported to: ###\n\t\t{}"
.format(str(train_name), trade_log_file_name + '.txt'))
log_file_readable.close()
log_file_raw = open(trade_log_file_name + '.json', 'w+')
json.dump(TI_end.trade_log, log_file_raw, indent=4)
print(
"### RAW trade_log of {} successfully exported to: ###\n\t\t{}"
.format(str(train_name),
trade_log_file_name + '.json'))
log_file_raw.close()
break
step += 1
TI_train = TI.Trade_Interface(config_account_name, config_currency_balance,
config_from, config_to, config_interval,
config_output_arena_csv,
config_output_raw_csv)
env = FX(TI_train,
_base_currency=config_base_currency,
_n_features=config_n_features,
_anita_switch=config_anita_switch)
DQN = DQN(len(env.TI_initial.currency_pairs),
env.n_actions,
env.n_features,
learning_rate=config_learning_rate,
reward_decay=config_reward_decay,
e_greedy=config_e_greedy,
replace_target_iter=config_replace_target_iter,
memory_size=config_memory_size,
output_graph=config_output_graph)
run_model(_train_episode=config_train_episode,
_learn_threshold=config_learn_threshold,
_learn_interval=config_learn_interval,
_base_currency=config_base_currency,
_trade_log_mode=config_trade_log_mode,
_trade_log_raw=config_trade_log_raw,
_trade_log_to_file=config_trade_log_to_file,
_show_checkout=config_show_checkout,
_show_step=config_show_step)
rate_reset = True
env.change_input()
state_episode = 0
RL.change_input_count()
print('change input:', RL.change_input)
else:
rate_reset = False
check_success_episode = 0
success_rate = 0
break
if action_step > 50:
break
step += 1
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
# maze game
env = Maze()
RL = DQN(n_actions=env.n_actions,
n_features=env.n_features,
n_flights=env.n_flights,
action_space=env.action_space)
env.after(100, run_maze)
env.mainloop()
reward = log2_shaping(reward, divide=1)
RL.store_transition(state, action, reward, next_state)
state = next_state
if ifrender:
env.render_img()
if RL.buffer.memory_counter > RL.memory_capacity:
RL.learn()
if done:
print("Epoch: {}/{}, highest: {}".format(
i, episodes, info['highest']))
scores.append(info['highest'])
if i % epsilon_decay_interval == 0:
RL.epsilon_decay(i, episodes)
break
return scores
if __name__ == "__main__":
RL = DQN(num_state=16, num_action=4, dueling=True)
scores = train(RL)
plt.figure(figsize=(18, 6), dpi=200)
plt.figure(1)
plt.plot(np.array(scores), c='r')
plt.ylabel('highest score')
plt.xlabel('training steps')
plt.savefig('result.jpg')
from maze_env import Maze
from RL_brain import DQN
import time
# Parameters
EPSILON = 0.9 # Greedy Policy
ALPHA = 0.1 # Learing Rate
LAMBDA = 0.9 # Discount Factor
MAX_EPISODE = 50
MAZE_SIZE = 5
TRAP_SET = [[0, 1], [2, 2], [3, 3], [3, 0]]
TREASURE_SET = [[4, 4]]
env = Maze(MAZE_SIZE)
RL = DQN(MAZE_SIZE, EPSILON, LAMBDA, ALPHA)
def update():
for episode in range(MAX_EPISODE):
O = env.reset() # return initial observation
step_number = 0
env.set_trap(TRAP_SET)
env.set_treasure(TREASURE_SET)
is_terminated = False
env.render()
while not is_terminated:
A = RL.choose_action(O)
# print(O)
OO, R, is_terminated = env.step(A)
# print(O,OO)
import gym
from RL_brain import DQN
env = gym.make('CartPole-v0')
env = env.unwrapped
print(env.action_space)
print(env.observation_space)
print(env.observation_space.high)
print(env.observation_space.low)
RL = DQN(n_actions=env.action_space.n,
n_features=env.observation_space.shape[0],
learning_rate=0.01, e_greedy=0.9,
replace_target_iter=100, memory_size=2000,
e_greedy_increment=0.001,)
total_steps = 0
for i_episode in range(100):
observation = env.reset()
ep_r = 0
while True:
env.render()
action = RL.choose_action(observation)
from RL_brain import DQN
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
env = gym.make('Pendulum-v0')
env = env.unwrapped
env.seed(1)
MEMORY_SIZE = 3000
ACTION_SPACE = 25
sess = tf.Session()
with tf.variable_scope('Natural_DQN'):
natural_DQN = DQN(n_actions=ACTION_SPACE,
n_features=3,
memory_size=MEMORY_SIZE,
e_greedy_increment=0.001,
dueling=False,
sess=sess)
with tf.variable_scope('Double_DQN'):
double_DQN = DQN(n_actions=ACTION_SPACE,
n_features=3,
memory_size=MEMORY_SIZE,
e_greedy_increment=0.001,
dueling=True,
sess=sess,
output_graph=True)
sess.run(tf.global_variables_initializer())
observation_, reward, done = env.step(action)
RL.memorize(observation, action, reward, observation_)
if (step > 200) and (step % 3 == 0):
RL.learn()
# swap observation
observation = observation_
# break while loop when end of this episode
if done:
break
step += 1
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
# maze game
env = Maze()
RL = DQN(env.n_actions,
memory_size=300,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=100)
env.after(100, run_maze)
env.mainloop()
