强化学习概述

import gymnasium as gym import numpy as np import json # 创建FrozenLake环境 env = gym.make('FrozenLake-v1', desc=None, map_name="4x4", is_slippery=True) # 环境基本信息 print(f"观察空间: {env.observation_space}") print(f"动作空间: {env.action_space}") print(f"动作含义: 0=左 1=下 2=右 3=上") # 随机策略运行 episodes = 1000 rewards = [] steps_list = [] success = 0 for ep in range(episodes): state, _ = env.reset(seed=ep) total_reward = 0 step = 0 done = False while not done: action = env.action_space.sample() # 随机策略 state, reward, terminated, truncated, info = env.step(action) total_reward += reward step += 1 done = terminated or truncated rewards.append(total_reward) steps_list.append(step) if total_reward > 0: success += 1 print(f"\\n随机策略统计 ({episodes}回合):") print(f"成功率: {success/episodes*100:.1f}%") print(f"平均步数: {np.mean(steps_list):.1f}") print(f"平均奖励: {np.mean(rewards):.4f}") print(f"奖励标准差: {np.std(rewards):.4f}") # 保存结果 result = { "episodes": episodes, "success_rate": round(success/episodes*100, 1), "avg_steps": round(np.mean(steps_list), 1), "avg_reward": round(float(np.mean(rewards)), 4), "success": success } with open("/var/www/ttl/rl/lesson01_result.json", "w") as f: json.dump(result, f) print("\\n✅验证通过 - 随机策略在FrozenLake上成功运行") env.close() # ============================================ # 扩展实验：不同地图配置对比 # ============================================ print("\n=== 扩展实验 ===") for map_name in ["4x4", "8x8"]: env2 = gym.make('FrozenLake-v1', map_name=map_name, is_slippery=True) wins = 0 for ep in range(5000): s, _ = env2.reset(seed=ep) done = False while not done: a = env2.action_space.sample() s, r, t, tr, _ = env2.step(a) done = t or tr if r > 0: wins += 1 print(f"{map_name}随机策略成功率: {wins/5000*100:.2f}%") # slippery vs non-slippery env3 = gym.make('FrozenLake-v1', map_name="4x4", is_slippery=False) wins = 0 for ep in range(5000): s, _ = env3.reset(seed=ep) done = False while not done: a = env3.action_space.sample() s, r, t, tr, _ = env3.step(a) done = t or tr if r > 0: wins += 1 print(f"4x4非滑行随机策略成功率: {wins/5000*100:.2f}%") print("扩展实验完成 - ✅")

输入: 环境env, 回合数N 输出: 平均奖励avg_reward 1. 初始化: total_reward = 0 2. FOR episode = 1 TO N: 3. state = env.reset() 4. done = False 5. WHILE NOT done: 6. action = 随机选择(env.action_space) 7. state, reward, done = env.step(action) 8. total_reward += reward 9. END WHILE 10. END FOR 11. avg_reward = total_reward / N 12. RETURN avg_reward