def see_path():
'''
show the path from q_table
:return:
'''
q_table = read_table()
env = Maze()
agent = DQN_Agent(actions=list(range(env.n_actions)))
agent.load_q_table(q_table)
s = env.reset()
while True:
env.render()
a = agent.get_path(s)
# print(a)
time.sleep(0.2)
s_, r, done = env.step(a)
s = s_
if done:
env.render()
break
def common_check(episodes=400):
'''
an ordinary learning process, and store q_table
:return:
'''
env = Maze()
agent = DQN_Agent(actions=list(range(env.n_actions)))
for episode in range(episodes):
s = env.reset()
episode_reward = 0
while True:
#env.render() # You can comment all render() to turn off the graphical interface in training process to accelerate your code.
a = agent.choose_action(s)
s_, r, done = env.step(a)
q_table = agent.update_q(s, s_, a, r)
episode_reward += r
s = s_
if done:
#env.render()
break
print('episode:', episode, 'episode_reward:', episode_reward)
store_table(q_table)
def check_converge_time():
'''
to show how many episodes needed to find the optimal path
for the first time
:return:
'''
env = Maze()
cvg_time = 0
for i in range(100):
print(i)
agent = DQN_Agent(actions=list(range(env.n_actions)))
flag = 0
for episode in range(300):
if flag:
break
s = env.reset()
episode_reward = 0
while True:
#env.render() # You can comment all render() to turn off the graphical interface in training process to accelerate your code.
a = agent.choose_action(s)
s_, r, done = env.step(a)
q_table = agent.update_q(s, s_, a, r)
episode_reward += r
if episode_reward == 4:
cvg_time += episode
flag = 1
s = s_
#print(s)
if done:
#env.render()
#time.sleep(0.5)
break
# print('episode:', episode, 'episode_reward:', episode_reward)
if flag == 0:
cvg_time += 300
print(cvg_time / 100)
episodes = 100
model_based_episodes = 5
env = Maze()
model = Model(actions=list(range(env.n_actions)))
agent = Agent(actions=list(range(
env.n_actions))) # 从range(4)，也就是0,1,2,3（上下右左）四个行为中选择
for episode in range(episodes): # 对于每一段，从开始到结束
s = env.reset()
episode_reward = 0
while True:
#env.render() # You can comment all render() to turn off the graphical interface in training process to accelerate your code.
# move one step
a = agent.choose_action(str(s))
s_, r, done = env.step(a)
# update Q model-free
agent.learn(str(s), a, r, str(s_), done)
model.store_transition(str(s), a, r, s_)
# update Q model-based
for n in range(model_based_episodes):
ss, sa = model.sample_s_a()
sr, ss_ = model.get_r_s_(ss, sa)
agent.learn(ss, sa, sr, str(ss_), done)
episode_reward += r
s = s_
if __name__ == "__main__":
### START CODE HERE ###
# This is an agent with random policy. You can learn how to interact with the environment through the code below.
# Then you can delete it and write your own code.
env = Maze()
training_epoch = 100 if maze == '1' else 1000
agent = D_Q_Agent(training_epoch)
for episode in range(training_epoch):
agent.if_rewarded = False
s = env.reset()
while True:
# env.render() # You can comment all render() to turn off the graphical interface in training process to accelerate your code.
chosen_direction = agent.choose_action(s, episode)
s_, r, done = env.step(chosen_direction)
agent.update_Q_value(s, chosen_direction, r)
if s_[-1]:
agent.if_rewarded = True
agent.if_rewarded_in_the_whole_training = True
s = s_
agent.simulative_training(100)
if done:
#env.render()
time.sleep(0.5)
break
print('episode:', episode)
print('Training Finished! Now Demonstrate the Optimal Policy:')
while True: