增強式學習

 1: import gymnasium as gym
 2: 
 3: # 建立執行環境
 4: env = gym.make('CartPole-v1')
 5: # 將執行環境初始化（回傳初始觀測值和額外資訊）
 6: observation, info = env.reset()
 7: for _ in range(100):
 8:     # 隨機從action_space中挑選下一動作(action)丟入step執行
 9:     observation, reward, terminated, truncated, info = env.step(env.action_space.sample())
10:     if terminated or truncated:
11:         break
12: env.close()
13: print("隨機亂動，遊戲結束！")

 1: import gymnasium as gym
 2: 
 3: env = gym.make('CartPole-v1')
 4: observation, info = env.reset()
 5: rewards = 0
 6: for _ in range(100):
 7:     action = env.action_space.sample()  # 仍然隨機產生 action
 8:     # action丟入環境執行，傳回環境資訊（5個回傳值）
 9:     observation, reward, terminated, truncated, info = env.step(action)
10:     rewards += reward
11:     print(observation)
12:     if terminated or truncated:  # 回合結束，可能柱子太傾斜或車子跑遠
13:         print("Rewards:", rewards)
14:         break
15: env.close()

1: import gymnasium as gym
2: env = gym.make('CartPole-v1')
3: print("動作空間:", env.action_space)
4: print("觀測空間:", env.observation_space)

 1: import numpy as np
 2: import gymnasium as gym
 3: import matplotlib.pyplot as plt
 4: 
 5: def run_one_episode(env, weights):
 6:     """用指定的權重跑一回合，回傳總 reward"""
 7:     observation, info = env.reset()
 8:     rewards = 0
 9:     for t in range(500):
10:         # 加權和決定動作
11:         action = 1 if np.dot(weights[:4], observation) + weights[4] >= 0 else 0
12:         observation, reward, terminated, truncated, info = env.step(action)
13:         rewards += reward
14:         if terminated or truncated:
15:             break
16:     return rewards
17: 
18: # 爬山演算法主程式
19: env = gym.make("CartPole-v1")
20: np.random.seed(10)
21: 
22: best_reward = 0
23: best_weights = np.random.rand(5)  # 初始權重隨機
24: reward_history = []
25: 
26: for i in range(300):
27:     # 在最佳權重上加入微小隨機擾動
28:     new_weights = best_weights + np.random.normal(0, 0.1, 5)
29:     r = run_one_episode(env, new_weights)
30:     reward_history.append(r)
31:     if r > best_reward:
32:         best_reward = r
33:         best_weights = new_weights
34:     if best_reward >= 500:  # CartPole-v1 滿分 500
35:         print(f"第 {i} 次迭代就找到滿分策略！")
36:         break
37: 
38: print(f"最佳 reward: {best_reward}")
39: env.close()
40: 
41: # 畫出學習過程
42: plt.figure(figsize=(10, 4))
43: plt.plot(reward_history)
44: plt.xlabel("Iteration")
45: plt.ylabel("Reward")
46: plt.title("Hill Climbing: Reward per Iteration")
47: plt.grid(True)
48: plt.tight_layout()
49: plt.savefig("images/hill_climbing_result.png", dpi=150)

 1: import gymnasium as gym
 2: import matplotlib.pyplot as plt
 3: import numpy as np
 4: 
 5: def run_episodes(strategy, n_episodes=200):
 6:     """跑 n_episodes 回合，回傳每回合的 reward 列表"""
 7:     env = gym.make('CartPole-v1')
 8:     rewards_list = []
 9:     for _ in range(n_episodes):
10:         obs, info = env.reset()
11:         total_reward = 0
12:         for t in range(500):
13:             if strategy == 'random':
14:                 action = env.action_space.sample()
15:             elif strategy == 'intuition':
16:                 # 直覺策略：桿子往左傾就往左推
17:                 action = 0 if obs[2] < 0 else 1
18:             obs, reward, terminated, truncated, info = env.step(action)
19:             total_reward += reward
20:             if terminated or truncated:
21:                 break
22:         rewards_list.append(total_reward)
23:     env.close()
24:     return rewards_list
25: 
26: # 只跑隨機策略
27: random_rewards = run_episodes('random')
28: print(f"隨機策略平均 reward: {np.mean(random_rewards):.1f}")

 1: import numpy as np
 2: import gymnasium as gym
 3: 
 4: def hill_climb(step_size, max_iter=300, seed=42):
 5:     """用指定步幅跑爬山演算法，回傳找到的最佳 reward 和歷史紀錄"""
 6:     env = gym.make('CartPole-v1')
 7:     np.random.seed(seed)
 8:     best_reward = 0
 9:     best_weights = np.random.rand(5)
10:     history = []
11: 
12:     for i in range(max_iter):
13:         new_weights = best_weights + np.random.normal(0, step_size, 5)
14:         obs, info = env.reset()
15:         total = 0
16:         for _ in range(500):
17:             action = 1 if np.dot(new_weights[:4], obs) + new_weights[4] >= 0 else 0
18:             obs, reward, terminated, truncated, info = env.step(action)
19:             total += reward
20:             if terminated or truncated:
21:                 break
22:         history.append(total)
23:         if total > best_reward:
24:             best_reward = total
25:             best_weights = new_weights
26:     env.close()
27:     return best_reward, history
28: 
29: best, hist = hill_climb(step_size=0.1)
30: print(f"步幅 0.1 → 最佳 reward: {best}")

 1: import gymnasium as gym
 2: import numpy as np
 3: 
 4: def my_strategy(obs):
 5:     """直覺策略：只看桿子角度"""
 6:     pos, vel, angle, ang_vel = obs
 7:     if angle < 0:
 8:         return 0  # 往左推
 9:     else:
10:         return 1  # 往右推
11: 
12: # 測試策略
13: env = gym.make('CartPole-v1')
14: rewards = []
15: for _ in range(100):
16:     obs, info = env.reset()
17:     total = 0
18:     for _ in range(500):
19:         action = my_strategy(obs)
20:         obs, reward, terminated, truncated, info = env.step(action)
21:         total += reward
22:         if terminated or truncated:
23:             break
24:     rewards.append(total)
25: env.close()
26: print(f"直覺策略平均 reward: {np.mean(rewards):.1f}")