二级建造师证怎么考青岛网站优化

引子

1. 古之成大事者，规模远大与综理密微，二者缺一不可。
2. 不管天气好坏，坚持每天前进大概30公里。
3. 起初店里的生意比较惨淡，他们有大把的时间来编写程序。

起因， 目的:

8x8的地图很容易失败， 这个问题，我之前讲过。

如何解决， 一句话，根据距离来修改奖励.

过程:

1. 先让 ChatGPT 推荐几种方法

聊天记录

• ChatGPT 推荐了5种方法，我试了， 都不行。失败率还是很高。
• 再推荐几种方法，还是不行。比如 DQN, 我试了，失败。机器人在某个地方左右反复移动，就是不前进。

2. 因此我建议，使用距离公式。

因为，我之前在什么地方看过类似的思路。具体想不起来了。可能是算法题里面见过的。
然后让 GPT 根据这个思路来写代码。
运行一下， ok!

import gymnasium as gym
import numpy as np
import matplotlib.pyplot as plt
import pickledef get_distance(state, goal_state, grid_size=8):"""计算当前位置到终点的曼哈顿距离"""x1, y1 = divmod(state, grid_size)x2, y2 = divmod(goal_state, grid_size)return abs(x1 - x2) + abs(y1 - y2)def run(episodes, is_training=True, render=False):env = gym.make('FrozenLake-v1', map_name="8x8", is_slippery=False)q = np.zeros((env.observation_space.n, env.action_space.n)) # init a 64 x 4 arraygoal_state = 63  # 8x8 终点在右下角（编号 63）lr = 0.9 # alpha or learning ratediscount_factor_g = 0.9 # gamma or discount rate. Near 0: more weight/reward placed on immediate state. Near 1: more on future state.epsilon = 1         # 1 = 100% random actionsepsilon_decay_rate = 0.0001        # epsilon decay rate. 1/0.0001 = 10,000rng = np.random.default_rng()   # random number generatorrewards_per_episode = np.zeros(episodes)for i in range(episodes):state = env.reset()[0]  # states: 0 to 63, 0=top left corner,63=bottom right cornerterminated = False      # True when fall in hole or reached goaltruncated = False       # True when actions > 200while not terminated and not truncated:if is_training and rng.random() < epsilon:action = env.action_space.sample() # actions: 0=left,1=down,2=right,3=upelse:action = np.argmax(q[state,:])new_state,reward,terminated,truncated,_ = env.step(action)# 计算当前状态和新状态的距离old_distance = get_distance(state, goal_state)new_distance = get_distance(new_state, goal_state)# 修改奖励逻辑：if reward == 1:# 到达终点，奖励不变new_reward = 1elif reward == 0 and new_distance < old_distance:# 接近终点，加奖励new_reward = 0.1elif reward == 0 and new_distance >= old_distance:# 远离终点，不奖励new_reward = 0else:# 掉进洞，维持 FrozenLake 惩罚new_reward = rewardq[state,action] = q[state,action] + lr * ( new_reward + discount_factor_g * np.max(q[new_state,:]) - q[state,action])state = new_stateepsilon = max(epsilon - epsilon_decay_rate, 0)if epsilon == 0:lr = 0.0001if reward == 1:rewards_per_episode[i] = 1env.close()plt.figure(figsize=(10, 8))sum_rewards = np.zeros(episodes)for t in range(episodes):sum_rewards[t] = np.sum(rewards_per_episode[max(0, t-100):(t+1)])plt.plot(sum_rewards)plt.savefig('frozen_lake8x8-distance-reward--1.png')return np.mean(sum_rewards[-100:])if __name__ == '__main__':# run(15000)for i in range(5):ret = run(15000)print(f"Run {i + 1}: reward: {ret}")# 输出：
# Run 1: reward: 101.0
# Run 2: reward: 101.0
# Run 3: reward: 101.0
# Run 4: reward: 101.0
# Run 5: reward: 101.0

3. 最近说一下这样做的理由

gpt 说的也挺好的:

📌 这样做的优点

✅ 更稳定的学习过程：机器人不会盲目探索，而是朝目标前进。
✅ 减少掉坑的次数：鼓励它更早学习正确路径。
✅ 适用于 8×8 复杂地图：比完全随机奖励的 Q-learning 更快收敛。

就是说，鼓励探索，探索的过程，也是有奖励的。

结论 + todo

最后的最后，分享一句话，来自 sentex：

RL 的难点，在于环境，奖励，观测。

---