def main():
# 初始化 环境
# 冰湖环境
# env = gym.make("FrozenLake-v0", is_slippery=False) # 0 left, 1 down, 2 right, 3 up
# env = FrozenLakeWapper(env)
# 悬崖环境
env = gym.make("CliffWalking-v0")
env = CliffWalkingWapper(env)
# 初始化 Agent
agent = SarsaAgent(obs_n=env.observation_space.n,
act_n=env.action_space.n,
learning_rate=0.1,
gamma=0.9,
e_greed=0.1)
# 开始训练
render = False
for episode in range(500):
ep_steps, ep_reward = run_episode(env, agent, render)
print('Episode %s: steps = %s , reward = %.1f' %
(episode, ep_steps, ep_reward))
# 每隔 20 个 episode 看一下效果
if episode % 20 == 0:
render = True
else:
render = False
# 训练结束，看一下效果
test_episode(env, agent)
def main():
# env = gym.make("FrozenLake-v0", is_slippery=False) # 0 left, 1 down, 2 right, 3 up
# env = FrozenLakeWapper(env)
env = gym.make("CliffWalking-v0") # 0 up, 1 right, 2 down, 3 left
env = CliffWalkingWapper(env)
agent = SarsaAgent(obs_n=env.observation_space.n,
act_n=env.action_space.n,
learning_rate=0.1,
gamma=0.9,
e_greed=0.1)
is_render = False
for episode in range(500):
ep_reward, ep_steps = run_episode(env, agent, is_render)
print('Episode %s: steps = %s , reward = %.1f' %
(episode, ep_steps, ep_reward))
# 每隔20个episode渲染一下看看效果
if episode % 20 == 0:
is_render = True
else:
is_render = False
# 训练结束，查看算法效果
test_episode(env, agent)
def train(cfg):
# env = gym.make("FrozenLake-v0", is_slippery=False) # 0 left, 1 down, 2 right, 3 up
# env = FrozenLakeWapper(env)
env = gym.make("CliffWalking-v0") # 0 up, 1 right, 2 down, 3 left
env = CliffWalkingWapper(env)
agent = QLearning(obs_dim=env.observation_space.n,
action_dim=env.action_space.n,
learning_rate=cfg.policy_lr,
gamma=cfg.gamma,
epsilon_start=cfg.epsilon_start,
epsilon_end=cfg.epsilon_end,
epsilon_decay=cfg.epsilon_decay)
render = False # 是否打开GUI画面
rewards = [] # 记录所有episode的reward
MA_rewards = [] # 记录滑动平均的reward
steps = [] # 记录所有episode的steps
for i_episode in range(1, cfg.max_episodes + 1):
ep_reward = 0 # 记录每个episode的reward
ep_steps = 0 # 记录每个episode走了多少step
obs = env.reset() # 重置环境, 重新开一局（即开始新的一个episode）
while True:
action = agent.sample(obs) # 根据算法选择一个动作
next_obs, reward, done, _ = env.step(action) # 与环境进行一个交互
# 训练 Q-learning算法
agent.learn(obs, action, reward, next_obs, done) # 不需要下一步的action
obs = next_obs # 存储上一个观察值
ep_reward += reward
ep_steps += 1 # 计算step数
if render:
env.render() #渲染新的一帧图形
if done:
break
steps.append(ep_steps)
rewards.append(ep_reward)
# 计算滑动平均的reward
if i_episode == 1:
MA_rewards.append(ep_reward)
else:
MA_rewards.append(0.9 * MA_rewards[-1] + 0.1 * ep_reward)
print('Episode %s: steps = %s , reward = %.1f, explore = %.2f' %
(i_episode, ep_steps, ep_reward, agent.epsilon))
# 每隔20个episode渲染一下看看效果
if i_episode % 20 == 0:
render = True
else:
render = False
agent.save() # 训练结束，保存模型
output_path = os.path.dirname(__file__) + "/result/"
# 检测是否存在文件夹
if not os.path.exists(output_path):
os.mkdir(output_path)
np.save(output_path + "rewards_train.npy", rewards)
np.save(output_path + "MA_rewards_train.npy", MA_rewards)
np.save(output_path + "steps_train.npy", steps)
def test(cfg):
env = gym.make("CliffWalking-v0") # 0 up, 1 right, 2 down, 3 left
env = CliffWalkingWapper(env)
agent = QLearning(obs_dim=env.observation_space.n,
action_dim=env.action_space.n,
learning_rate=cfg.policy_lr,
gamma=cfg.gamma,
epsilon_start=cfg.epsilon_start,
epsilon_end=cfg.epsilon_end,
epsilon_decay=cfg.epsilon_decay)
agent.load() # 导入保存的模型
rewards = [] # 记录所有episode的reward
MA_rewards = [] # 记录滑动平均的reward
steps = [] # 记录所有episode的steps
for i_episode in range(1, 10 + 1):
ep_reward = 0 # 记录每个episode的reward
ep_steps = 0 # 记录每个episode走了多少step
obs = env.reset() # 重置环境, 重新开一局（即开始新的一个episode）
while True:
action = agent.predict(obs) # 根据算法选择一个动作
next_obs, reward, done, _ = env.step(action) # 与环境进行一个交互
obs = next_obs # 存储上一个观察值
time.sleep(0.5)
env.render()
ep_reward += reward
ep_steps += 1 # 计算step数
if done:
break
steps.append(ep_steps)
rewards.append(ep_reward)
# 计算滑动平均的reward
if i_episode == 1:
MA_rewards.append(ep_reward)
else:
MA_rewards.append(0.9 * MA_rewards[-1] + 0.1 * ep_reward)
print('Episode %s: steps = %s , reward = %.1f' %
(i_episode, ep_steps, ep_reward))
plt.plot(MA_rewards)
plt.show()
action = np.random.choice(action_list) # 最大值随机选取
return action
def learn(self, state, action, reward, next_state, next_action, done):
predict_Q = self.Q_TABLE[state, action] # Q_table表找到对应Q的评估值
if done:
target_Q = reward
else:
# Q <- Q + a*[(R + y*next_Q) - Q] Q_TABLE预测 max_action(Q(s',a))
target_Q = reward + GAMMA * np.max(self.Q_TABLE[next_state,:])
self.Q_TABLE[state, action] = self.Q_TABLE[state, action] + LEARNING_RATE * (target_Q - predict_Q)
if __name__ == '__main__':
env = gym.make("CliffWalking-v0")
env = CliffWalkingWapper(env)
dim_state = env.observation_space.n # 48
dim_action = env.action_space.n # 4
agent = QLearningAgent(dim_state, dim_action)
for epoch in range(500):
state = env.reset() # 开始一局游戏
total_rewards, total_steps = 0, 0
action = agent.choose_action(state)
while True:
#env.render()
next_state, reward, done, _ = env.step(action) # 采取该行为获取下一个state 及分数
next_action = agent.choose_action(next_state) # 行为 概率
agent.learn(state, action, reward, next_state, next_action, done)
action = next_action