{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# SARSA 与 Q-learning 在 CliffWalking-v1（悬崖行走）上的对比。\n",
    "\n",
    "\n",
    "\n",
    "配套教程：[Q-learning 与 SARSA](https://dqn.cn/basics/q-learning-sarsa)（DQN.CN 强化学习中文教程）\n",
    "\n",
    "源码：`examples/basics/cliff_q_sarsa.py`，本 notebook 由 `scripts/build-notebooks.mjs` 自动生成，代码与 .py 文件一致。代码默认在 CPU 上运行，固定了随机种子；部分脚本支持设置环境变量 `DQN_QUICK=1` 只跑很短的版本（见下一个单元）。"
   ],
   "id": "basics-cliff_q_sarsa-py-0"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 环境准备：在 Colab 等缺少依赖的环境中自动安装（本地已装好依赖时什么也不做）\n",
    "import importlib.util, os, subprocess, sys\n",
    "_need = [m for m in [\"gymnasium\"] if importlib.util.find_spec(m) is None]\n",
    "if _need or \"google.colab\" in sys.modules:\n",
    "    subprocess.run([sys.executable, \"-m\", \"pip\", \"install\", \"-q\", \"gymnasium>=1.0\"], check=True)\n",
    "sys.argv = sys.argv[:1]   # notebook 内核会带上 -f 等参数，清掉以免脚本把它们当成命令行参数\n",
    "# os.environ[\"DQN_QUICK\"] = \"1\"   # 取消注释：支持该开关的脚本只跑很短的版本，用于快速检查"
   ],
   "id": "basics-cliff_q_sarsa-py-1"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import gymnasium as gym\n",
    "import numpy as np\n",
    "\n",
    "env = gym.make(\"CliffWalking-v1\")\n",
    "nS, nA = env.observation_space.n, env.action_space.n"
   ],
   "id": "basics-cliff_q_sarsa-py-2"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def eps_greedy(Q, s, eps, rng):\n",
    "    if rng.random() < eps:\n",
    "        return int(rng.integers(nA))\n",
    "    q = Q[s]\n",
    "    return int(rng.choice(np.flatnonzero(q == q.max())))   # 随机打破平局"
   ],
   "id": "basics-cliff_q_sarsa-py-3"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train(method, episodes=500, alpha=0.1, gamma=1.0, eps=0.1, seed=0):\n",
    "    rng = np.random.default_rng(seed)\n",
    "    Q = np.zeros((nS, nA))\n",
    "    returns = []\n",
    "    for ep in range(episodes):\n",
    "        s, _ = env.reset(seed=seed * 10_000 + ep)\n",
    "        a = eps_greedy(Q, s, eps, rng)\n",
    "        G, done = 0.0, False\n",
    "        while not done:\n",
    "            s2, r, term, trunc, _ = env.step(a)\n",
    "            done = term or trunc\n",
    "            G += r\n",
    "            a2 = eps_greedy(Q, s2, eps, rng)\n",
    "            if method == \"sarsa\":\n",
    "                target = r + gamma * Q[s2, a2] * (not term)\n",
    "            else:  # Q-learning\n",
    "                target = r + gamma * np.max(Q[s2]) * (not term)\n",
    "            Q[s, a] += alpha * (target - Q[s, a])\n",
    "            s, a = s2, a2\n",
    "        returns.append(G)\n",
    "    return Q, np.array(returns)"
   ],
   "id": "basics-cliff_q_sarsa-py-4"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def greedy_return(Q):\n",
    "    s, _ = env.reset()\n",
    "    G = 0\n",
    "    for _ in range(100):\n",
    "        s, r, term, trunc, _ = env.step(int(np.argmax(Q[s])))\n",
    "        G += r\n",
    "        if term or trunc:\n",
    "            break\n",
    "    return G"
   ],
   "id": "basics-cliff_q_sarsa-py-5"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for method in [\"sarsa\", \"q_learning\"]:\n",
    "    online, greedy = [], []\n",
    "    for seed in range(10):\n",
    "        Q, rets = train(method, seed=seed)\n",
    "        online.append(rets[-100:].mean())\n",
    "        greedy.append(greedy_return(Q))\n",
    "    print(f\"{method:10s} 训练中（ε=0.1）最后100回合平均回报: {np.mean(online):7.1f} | \"\n",
    "          f\"贪心策略回报: {np.mean(greedy):6.1f}\")"
   ],
   "id": "basics-cliff_q_sarsa-py-6"
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  },
  "colab": {
   "provenance": []
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}