def main():
model = PolicyGradient(
gym.make('CartPole-v1'),
n_in=4,
hidden_net=lambda x: tf.layers.dense(x, 10, tf.nn.elu),
n_out=2)
model.learn()
model.play()
def main():
env = grc.RemoteEnv('tmp/sock')
# Policy gradient has high variance, seed for reproducability
env.seed(1)
RENDER_ENV = False
rewards = []
INITIAL_EPSILON = 0.7
EPSILON_GREEDY_INCREMENT = 0.01
# Load checkpoint
load_version = "2018-06-05 18:24:13"
timestamp = strftime("%Y-%m-%d %H:%M:%S", gmtime())
load_path = "output/model2/{}/SonicTheHedgehog.ckpt".format(load_version)
PG = PolicyGradient(
n_x = [112,112,3], #env.observation_space.shape,
n_y = env.action_space.n,
learning_rate=0.02,
reward_decay=0.99,
load_path=load_path,
epsilon_max=0.98,
epsilon_greedy_increment=EPSILON_GREEDY_INCREMENT,
initial_epsilon = INITIAL_EPSILON
)
observation = env.reset()
# print("obs", observation)
episode_reward = 0
tic = time.clock()
while True:
if RENDER_ENV: env.render()
# 1. Choose an action based on observation
observation = observation[:,96:,:] # make square, keep right sight of image
observation = observation[::2,::2,:] # downsample to [112,112,3]
observation = observation / 255 # normalize
action = PG.choose_action(observation)
# 2. Take action in the environment
observation_, reward, done, info = env.step(action)
# 4. Store transition for training
PG.store_transition(observation, action, reward)
episode_rewards_sum = sum(PG.episode_rewards)
toc = time.clock()
elapsed_sec = toc - tic
# Save new observation
observation = observation_
if done:
episode_rewards_sum = sum(PG.episode_rewards)
rewards.append(episode_rewards_sum)
max_reward_so_far = np.amax(rewards)
if episode_rewards_sum == 0.0:
print("-----------------------------------")
print("Backtrack epsilon for more exploration...")
PG.epsilon = max(PG.epsilon - EPSILON_GREEDY_INCREMENT, INITIAL_EPSILON)
print("==========================================")
print("Epsilon: ", PG.epsilon)
print("Seconds: ", elapsed_sec)
print("Reward: ", episode_rewards_sum)
print("Max reward so far: ", max_reward_so_far)
# 5. Train neural network
tic = time.clock()
discounted_episode_rewards_norm = PG.learn()
toc = time.clock()
elapsed_sec = toc - tic
print("Train Seconds: ", elapsed_sec)
observation = env.reset()
def main():
model = PolicyGradient(gym.make('CartPole-v1'),
n_in=4, n_hidden=[10], n_out=2)
model.learn()
model.play()
import matplotlib.pyplot as plt
DISPLAY_REWARD_THRESHOLD = -200
RENDER = False
is_train = False
env = gym.make('MountainCar-v0')
env.seed(1)
env = env.unwrapped
print(env.action_space)
print(env.observation_space)
print(env.observation_space.high)
RL = PolicyGradient(n_actions=env.action_space.n,
n_features=env.observation_space.shape[0],
learning_rate=0.01,
reward_decay=0.98)
if not is_train:
model_file = RL.restore_file
RL.saver.restore(RL.sess, model_file)
max_reward = -200
for i_episode in range(1000):
observation = env.reset()
running_reward = 0
i = 0
while True:
if RENDER:
env.render()
def main():
model = PolicyGradient()
model.learn(gym.make("CartPole-v0"))
model.play(gym.make("CartPole-v0"))
# break while loop when end of this episode
if done:
break
step += 1
print(totalReward)
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
env = Plane()
# RL = QLearningTable(actions=list(range(env.n_actions)))
RL = DeepQNetwork(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
)
RL = PolicyGradient(env.n_actions, env.n_features,
learning_rate=0.01,
reward_decay=0.9,
# output_graph=True
)
env.after(100, update)
env.mainloop()
#DQN
RL.plot_cost()
# training, testing data split train_start = '2014-01-01' train_test_split = '2015-01-01' training_data = df[train_start:train_test_split] train_test_split = '2015-01-01' #train_test_split = '2016-01-01' test_end = '2016-01-01' testing_data = df[train_test_split:test_end] #%% train env = MarketEnv(training_data, 60) #pg = PolicyGradient(env, gamma= 0.9,file_name = "pg_3.h5") pg = PolicyGradient(env, gamma=0.9, file_name="new34.h5") pg.train() #%% test env_test = MarketEnv(testing_data, 60) pg_test = PolicyGradient(env_test, gamma=0.9, weights_path="new34.h5") model = pg_test.model act = [-1, 0, 1] env_test.reset() observation = env_test.reset() game_over = False inputs = [] outputs = [] predicteds = []
def main():
model = PolicyGradient(lambda x: tf.layers.dense(x, 4, tf.nn.relu))
model.learn(gym.make("CartPole-v0"))
model.play(gym.make("CartPole-v0"))
import gym
from gym import wrappers
import numpy as np
import matplotlib.pyplot as plt
from pg import PolicyGradient
DISPLAY_REWARD_THRESHOLD = -2000
RENDER = False
env = gym.make('MountainCar-v0')
env.seed(1)
pg = PolicyGradient(n_actions=env.action_space.n,
n_features=env.observation_space.shape[0],
learning_rate=0.02,
reward_decay=0.995)
"""
这里先用env.unwrapper来移除env上包裹的所有wrapper（其中一个让env的每个episode最大step设置为200），
然后给env包裹上Monitor Wrapper来实现渲染
"""
env = env.unwrapped
env = wrappers.Monitor(env, "../gym-results", force=True)
for episode in range(1000):
obs = env.reset()
while True:
if RENDER:
env.render()
action = pg.choose_action(obs)
obs_, reward, done, info = env.step(action)
# print("cur_reward = ", reward)
INITIAL_EPSILON = 0.7
EPSILON_GREEDY_INCREMENT = 0.01
if __name__ == "__main__":
# Load checkpoint
load_version = "2018-06-05 16:31:58"
timestamp = strftime("%Y-%m-%d %H:%M:%S", gmtime())
load_path = "output/model2/{}/SonicTheHedgehog.ckpt".format(load_version)
save_path = "output/model2/{}/SonicTheHedgehog.ckpt".format(timestamp)
PG = PolicyGradient(
n_x=[112, 112, 3], #env.observation_space.shape,
n_y=env.action_space.n,
learning_rate=0.02,
reward_decay=0.99,
load_path=load_path,
save_path=save_path,
epsilon_max=0.98,
epsilon_greedy_increment=EPSILON_GREEDY_INCREMENT,
initial_epsilon=INITIAL_EPSILON)
for episode in range(EPISODES):
observation = env.reset()
# print("obs", observation)
episode_reward = 0
tic = time.clock()
while True:
if RENDER_ENV: env.render()