observation = observation_
# 如果终止, 就跳出循环
if done:
steps.append(step)
episodes.append(episode)
break
step += 1 # 总步数
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
env = Maze()
RL = DoubleDQN(env.n_actions,
env.n_features,#observation/state 的属性，如长宽高
learning_rate=0.01,
reward_decay=0.9,
dueling=True,
e_greedy=0.9,
replace_target_iter=200, # 每 200 步替换一次 target_net 的参数
memory_size=2000, # 记忆上限
# output_graph=True # 是否输出 tensorboard 文件
)
env.after(100, run_maze)#进行强化学习训练
env.mainloop()
# 观看训练时间曲线
his_double = np.vstack((episodes, steps))
from agent import Agent, D_Q_Agent
import time
maze = '2'
if maze == '1':
from maze_env1 import Maze
elif maze == '2':
from maze_env2 import Maze
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_