{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Balancing immediate and long-term goals\n",
    "\n",
    "> 그로킹 심층 강화학습 중 3장 내용인 \"순간 목표와 장기 목표간의 균형\"에 대한 내용입니다.\n",
    "\n",
    "- hide: true\n",
    "- toc: true \n",
    "- badges: true\n",
    "- comments: true\n",
    "- author: Chanseok Kang\n",
    "- categories: [Python, Reinforcement_Learning, Grokking_Deep_Reinforcement_Learning]\n",
    "- permalink: /book/:title:output_ext\n",
    "- search_exclude: false"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "> Note: 라이브러리 설치를 위해 아래의 패키지들을 설치해주기 바랍니다."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#collapse\n",
    "!pip install tqdm numpy scikit-learn pyglet setuptools && \\\n",
    "!pip install gym asciinema pandas tabulate tornado==5.* PyBullet && \\\n",
    "!pip install git+https://github.com/pybox2d/pybox2d#egg=Box2D && \\\n",
    "!pip install git+https://github.com/mimoralea/gym-bandits#egg=gym-bandits && \\\n",
    "!pip install git+https://github.com/mimoralea/gym-walk#egg=gym-walk && \\\n",
    "!pip install git+https://github.com/mimoralea/gym-aima#egg=gym-aima && \\\n",
    "!pip install gym[atari]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import gym, gym_walk, gym_aima"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 정책 반복법, 가치 반복법\n",
    "\n",
    "**참고**: 해당 노트북에서 사용되는 환경에 대한 정보는 아래 링크를 참고하시기 바랍니다.\n",
    "\n",
    "- [gym_walk](https://github.com/mimoralea/gym-walk)\n",
    "- [gym_aima](https://github.com/mimoralea/gym-aima)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import warnings ; warnings.filterwarnings('ignore')\n",
    "\n",
    "import gym, gym_walk, gym_aima\n",
    "import numpy as np\n",
    "from pprint import pprint\n",
    "from tqdm import tqdm_notebook as tqdm\n",
    "\n",
    "from itertools import cycle\n",
    "\n",
    "import random\n",
    "\n",
    "np.set_printoptions(suppress=True)\n",
    "random.seed(123); np.random.seed(123)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 출력을 위한 helper function"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_policy(pi, P, action_symbols=('<', 'v', '>', '^'), n_cols=4, title='정책:'):\n",
    "    print(title)\n",
    "    arrs = {k:v for k,v in enumerate(action_symbols)}\n",
    "    for s in range(len(P)):\n",
    "        a = pi(s)\n",
    "        print(\"| \", end=\"\")\n",
    "        if np.all([done for action in P[s].values() for _, _, _, done in action]):\n",
    "            print(\"\".rjust(9), end=\" \")\n",
    "        else:\n",
    "            print(str(s).zfill(2), arrs[a].rjust(6), end=\" \")\n",
    "        if (s + 1) % n_cols == 0: print(\"|\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_state_value_function(V, P, n_cols=4, prec=3, title='상태-가치 함수:'):\n",
    "    print(title)\n",
    "    for s in range(len(P)):\n",
    "        v = V[s]\n",
    "        print(\"| \", end=\"\")\n",
    "        if np.all([done for action in P[s].values() for _, _, _, done in action]):\n",
    "            print(\"\".rjust(9), end=\" \")\n",
    "        else:\n",
    "            print(str(s).zfill(2), '{}'.format(np.round(v, prec)).rjust(6), end=\" \")\n",
    "        if (s + 1) % n_cols == 0: print(\"|\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_action_value_function(Q, \n",
    "                                optimal_Q=None, \n",
    "                                action_symbols=('<', '>'), \n",
    "                                prec=3, \n",
    "                                title='행동-가치 함수:'):\n",
    "    vf_types=('',) if optimal_Q is None else ('', '*', 'err')\n",
    "    headers = ['s',] + [' '.join(i) for i in list(itertools.product(vf_types, action_symbols))]\n",
    "    print(title)\n",
    "    states = np.arange(len(Q))[..., np.newaxis]\n",
    "    arr = np.hstack((states, np.round(Q, prec)))\n",
    "    if not (optimal_Q is None):\n",
    "        arr = np.hstack((arr, np.round(optimal_Q, prec), np.round(optimal_Q-Q, prec)))\n",
    "    print(tabulate(arr, headers, tablefmt=\"fancy_grid\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def probability_success(env, pi, goal_state, n_episodes=100, max_steps=200):\n",
    "    random.seed(123); np.random.seed(123) ; env.seed(123)\n",
    "    results = []\n",
    "    for _ in range(n_episodes):\n",
    "        state, done, steps = env.reset(), False, 0\n",
    "        while not done and steps < max_steps:\n",
    "            state, _, done, h = env.step(pi(state))\n",
    "            steps += 1\n",
    "        results.append(state == goal_state)\n",
    "    return np.sum(results)/len(results)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "def mean_return(env, pi, n_episodes=100, max_steps=200):\n",
    "    random.seed(123); np.random.seed(123) ; env.seed(123)\n",
    "    results = []\n",
    "    for _ in range(n_episodes):\n",
    "        state, done, steps = env.reset(), False, 0\n",
    "        results.append(0.0)\n",
    "        while not done and steps < max_steps:\n",
    "            state, reward, done, _ = env.step(pi(state))\n",
    "            results[-1] += reward\n",
    "            steps += 1\n",
    "    return np.mean(results)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Slippery Walk Five MDP and sample policy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "|           | 01      < | 02      < | 03      < | 04      < | 05      < |           |\n",
      "Reaches goal 7.00%. Obtains an average undiscounted return of 0.0700.\n"
     ]
    }
   ],
   "source": [
    "env = gym.make('SlipperyWalkFive-v0')\n",
    "P = env.env.P\n",
    "init_state = env.reset()\n",
    "goal_state = 6\n",
    "\n",
    "LEFT, RIGHT = range(2)\n",
    "pi = lambda s: {\n",
    "    0:LEFT, 1:LEFT, 2:LEFT, 3:LEFT, 4:LEFT, 5:LEFT, 6:LEFT\n",
    "}[s]\n",
    "print_policy(pi, P, action_symbols=('<', '>'), n_cols=7)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 정책 평가법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def policy_evaluation(pi, P, gamma=1.0, theta=1e-10):\n",
    "    prev_V = np.zeros(len(P), dtype=np.float64)\n",
    "    while True:\n",
    "        V = np.zeros(len(P), dtype=np.float64)\n",
    "        for s in range(len(P)):\n",
    "            for prob, next_state, reward, done in P[s][pi(s)]:\n",
    "                V[s] += prob * (reward + gamma * prev_V[next_state] * (not done))\n",
    "        if np.max(np.abs(prev_V - V)) < theta:\n",
    "            break\n",
    "        prev_V = V.copy()\n",
    "    return V"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "|           | 01 0.00275 | 02 0.01099 | 03 0.03571 | 04 0.10989 | 05 0.33242 |           |\n"
     ]
    }
   ],
   "source": [
    "V = policy_evaluation(pi, P)\n",
    "print_state_value_function(V, P, n_cols=7, prec=5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 정책 개선법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def policy_improvement(V, P, gamma=1.0):\n",
    "    Q = np.zeros((len(P), len(P[0])), dtype=np.float64)\n",
    "    for s in range(len(P)):\n",
    "        for a in range(len(P[s])):\n",
    "            for prob, next_state, reward, done in P[s][a]:\n",
    "                Q[s][a] += prob * (reward + gamma * V[next_state] * (not done))\n",
    "    new_pi = lambda s: {s:a for s, a in enumerate(np.argmax(Q, axis=1))}[s]\n",
    "    return new_pi"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "|           | 01      > | 02      > | 03      > | 04      > | 05      > |           |\n",
      "Reaches goal 93.00%. Obtains an average undiscounted return of 0.9300.\n"
     ]
    }
   ],
   "source": [
    "improved_pi = policy_improvement(V, P)\n",
    "print_policy(improved_pi, P, action_symbols=('<', '>'), n_cols=7)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, improved_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, improved_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "|           | 01 0.66758 | 02 0.89011 | 03 0.96429 | 04 0.98901 | 05 0.99725 |           |\n"
     ]
    }
   ],
   "source": [
    "# how about we evaluate the improved policy?\n",
    "improved_V = policy_evaluation(improved_pi, P)\n",
    "print_state_value_function(improved_V, P, n_cols=7, prec=5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "|           | 01      > | 02      > | 03      > | 04      > | 05      > |           |\n",
      "Reaches goal 93.00%. Obtains an average undiscounted return of 0.9300.\n"
     ]
    }
   ],
   "source": [
    "# can we improved the improved policy?\n",
    "improved_improved_pi = policy_improvement(improved_V, P)\n",
    "print_policy(improved_improved_pi, P, action_symbols=('<', '>'), n_cols=7)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, improved_improved_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, improved_improved_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "|           | 01 0.66758 | 02 0.89011 | 03 0.96429 | 04 0.98901 | 05 0.99725 |           |\n"
     ]
    }
   ],
   "source": [
    "# it is the same policy\n",
    "# if we evaluate again, we can see there is nothing to improve \n",
    "# that also means we reached the optimal policy\n",
    "improved_improved_V = policy_evaluation(improved_improved_pi, P)\n",
    "print_state_value_function(improved_improved_V, P, n_cols=7, prec=5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# state-value function didn't improve, then we reach the optimal policy\n",
    "assert np.all(improved_V == improved_improved_V)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 정책 반복법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "def policy_iteration(P, gamma=1.0, theta=1e-10):\n",
    "    random_actions = np.random.choice(tuple(P[0].keys()), len(P))\n",
    "    pi = lambda s: {s:a for s, a in enumerate(random_actions)}[s]\n",
    "    while True:\n",
    "        old_pi = {s:pi(s) for s in range(len(P))}\n",
    "        V = policy_evaluation(pi, P, gamma, theta)\n",
    "        pi = policy_improvement(V, P, gamma)\n",
    "        if old_pi == {s:pi(s) for s in range(len(P))}:\n",
    "            break\n",
    "    return V, pi"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "|           | 01      > | 02      > | 03      > | 04      > | 05      > |           |\n",
      "Reaches goal 93.00%. Obtains an average undiscounted return of 0.9300.\n",
      "\n",
      "상태-가치 함수:\n",
      "|           | 01 0.66758 | 02 0.89011 | 03 0.96429 | 04 0.98901 | 05 0.99725 |           |\n"
     ]
    }
   ],
   "source": [
    "optimal_V, optimal_pi = policy_iteration(P)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(optimal_pi, P, action_symbols=('<', '>'), n_cols=7)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, optimal_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, optimal_pi)))\n",
    "print()\n",
    "print_state_value_function(optimal_V, P, n_cols=7, prec=5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "assert np.all(improved_V == optimal_V)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Frozen Lake MDP and sample policies"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      > | 01      < | 02      v | 03      ^ |\n",
      "| 04      < |           | 06      > |           |\n",
      "| 08      ^ | 09      v | 10      ^ |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 12.00%. Obtains an average undiscounted return of 0.1200.\n"
     ]
    }
   ],
   "source": [
    "env = gym.make('FrozenLake-v0')\n",
    "P = env.env.P\n",
    "init_state = env.reset()\n",
    "goal_state = 15\n",
    "\n",
    "LEFT, DOWN, RIGHT, UP = range(4)\n",
    "random_pi = lambda s: {\n",
    "    0:RIGHT, 1:LEFT, 2:DOWN, 3:UP,\n",
    "    4:LEFT, 5:LEFT, 6:RIGHT, 7:LEFT,\n",
    "    8:UP, 9:DOWN, 10:UP, 11:LEFT,\n",
    "    12:LEFT, 13:RIGHT, 14:DOWN, 15:LEFT\n",
    "}[s]\n",
    "print_policy(random_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, random_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, random_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      > | 01      > | 02      v | 03      < |\n",
      "| 04      v |           | 06      v |           |\n",
      "| 08      > | 09      > | 10      v |           |\n",
      "|           | 13      > | 14      > |           |\n",
      "Reaches goal 5.00%. Obtains an average undiscounted return of 0.0500.\n"
     ]
    }
   ],
   "source": [
    "go_get_pi = lambda s: {\n",
    "    0:RIGHT, 1:RIGHT, 2:DOWN, 3:LEFT,\n",
    "    4:DOWN, 5:LEFT, 6:DOWN, 7:LEFT,\n",
    "    8:RIGHT, 9:RIGHT, 10:DOWN, 11:LEFT,\n",
    "    12:LEFT, 13:RIGHT, 14:RIGHT, 15:LEFT\n",
    "}[s]\n",
    "print_policy(go_get_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, go_get_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, go_get_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      ^ |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      > |           |\n",
      "Reaches goal 52.00%. Obtains an average undiscounted return of 0.5200.\n"
     ]
    }
   ],
   "source": [
    "careful_pi = lambda s: {\n",
    "    0:LEFT, 1:UP, 2:UP, 3:UP,\n",
    "    4:LEFT, 5:LEFT, 6:UP, 7:LEFT,\n",
    "    8:UP, 9:DOWN, 10:LEFT, 11:LEFT,\n",
    "    12:LEFT, 13:RIGHT, 14:RIGHT, 15:LEFT\n",
    "}[s]\n",
    "print_policy(careful_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, careful_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, careful_pi)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 정책 평가법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00 0.4079 | 01 0.3754 | 02 0.3543 | 03 0.3438 |\n",
      "| 04 0.4203 |           | 06 0.1169 |           |\n",
      "| 08 0.4454 | 09  0.484 | 10 0.4328 |           |\n",
      "|           | 13 0.5884 | 14 0.7107 |           |\n"
     ]
    }
   ],
   "source": [
    "V = policy_evaluation(careful_pi, P, gamma=0.99)\n",
    "print_state_value_function(V, P, prec=4)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 정책 개선법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n"
     ]
    }
   ],
   "source": [
    "careful_plus_pi = policy_improvement(V, P, gamma=0.99)\n",
    "print_policy(careful_plus_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, careful_plus_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, careful_plus_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00  0.542 | 01 0.4988 | 02 0.4707 | 03 0.4569 |\n",
      "| 04 0.5585 |           | 06 0.3583 |           |\n",
      "| 08 0.5918 | 09 0.6431 | 10 0.6152 |           |\n",
      "|           | 13 0.7417 | 14 0.8628 |           |\n"
     ]
    }
   ],
   "source": [
    "new_V = policy_evaluation(careful_plus_pi, P, gamma=0.99)\n",
    "print_state_value_function(new_V, P, prec=4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00 0.1341 | 01 0.1234 | 02 0.1164 | 03  0.113 |\n",
      "| 04 0.1381 |           | 06 0.2414 |           |\n",
      "| 08 0.1464 | 09 0.1591 | 10 0.1824 |           |\n",
      "|           | 13 0.1533 | 14 0.1521 |           |\n"
     ]
    }
   ],
   "source": [
    "print_state_value_function(new_V - V, P, prec=4)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Alternating between evaluation and improvement"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      ^ | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      ^ |           | 06      ^ |           |\n",
      "| 08      < | 09      < | 10      < |           |\n",
      "|           | 13      < | 14      < |           |\n",
      "Reaches goal 0.00%. Obtains an average undiscounted return of 0.0000.\n"
     ]
    }
   ],
   "source": [
    "adversarial_pi = lambda s: {\n",
    "    0:UP, 1:UP, 2:UP, 3:UP,\n",
    "    4:UP, 5:LEFT, 6:UP, 7:LEFT,\n",
    "    8:LEFT, 9:LEFT, 10:LEFT, 11:LEFT,\n",
    "    12:LEFT, 13:LEFT, 14:LEFT, 15:LEFT\n",
    "}[s]\n",
    "print_policy(adversarial_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, adversarial_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, adversarial_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00    0.0 | 01    0.0 | 02    0.0 | 03    0.0 |\n",
      "| 04    0.0 |           | 06    0.0 |           |\n",
      "| 08    0.0 | 09    0.0 | 10    0.0 |           |\n",
      "|           | 13    0.0 | 14    0.0 |           |\n"
     ]
    }
   ],
   "source": [
    "V = policy_evaluation(adversarial_pi, P, gamma=0.99)\n",
    "print_state_value_function(V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      < | 02      < | 03      < |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      < | 09      < | 10      < |           |\n",
      "|           | 13      < | 14      v |           |\n",
      "Reaches goal 0.00%. Obtains an average undiscounted return of 0.0000.\n"
     ]
    }
   ],
   "source": [
    "i_pi = policy_improvement(V, P, gamma=0.99)\n",
    "print_policy(i_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, i_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, i_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00    0.0 | 01    0.0 | 02   0.04 | 03   0.02 |\n",
      "| 04    0.0 |           | 06   0.07 |           |\n",
      "| 08    0.0 | 09    0.0 | 10   0.19 |           |\n",
      "|           | 13    0.0 | 14    0.5 |           |\n"
     ]
    }
   ],
   "source": [
    "i_V = policy_evaluation(i_pi, P, gamma=0.99)\n",
    "print_state_value_function(i_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      v | 02      > | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      < | 09      v | 10      < |           |\n",
      "|           | 13      v | 14      > |           |\n",
      "Reaches goal 0.00%. Obtains an average undiscounted return of 0.0000.\n"
     ]
    }
   ],
   "source": [
    "ii_pi = policy_improvement(i_V, P, gamma=0.99)\n",
    "print_policy(ii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, ii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, ii_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00    0.0 | 01   0.05 | 02   0.16 | 03   0.15 |\n",
      "| 04    0.0 |           | 06   0.17 |           |\n",
      "| 08    0.0 | 09   0.22 | 10   0.35 |           |\n",
      "|           | 13   0.33 | 14   0.67 |           |\n"
     ]
    }
   ],
   "source": [
    "ii_V = policy_evaluation(ii_pi, P, gamma=0.99)\n",
    "print_state_value_function(ii_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      v | 01      > | 02      > | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      v | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      > |           |\n",
      "Reaches goal 20.00%. Obtains an average undiscounted return of 0.2000.\n"
     ]
    }
   ],
   "source": [
    "iii_pi = policy_improvement(ii_V, P, gamma=0.99)\n",
    "print_policy(iii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, iii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, iii_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00   0.12 | 01   0.09 | 02   0.19 | 03   0.19 |\n",
      "| 04   0.15 |           | 06    0.2 |           |\n",
      "| 08   0.19 | 09   0.38 | 10   0.43 |           |\n",
      "|           | 13   0.53 | 14   0.71 |           |\n"
     ]
    }
   ],
   "source": [
    "iii_V = policy_evaluation(iii_pi, P, gamma=0.99)\n",
    "print_state_value_function(iii_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      > | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 73.00%. Obtains an average undiscounted return of 0.7300.\n"
     ]
    }
   ],
   "source": [
    "iiii_pi = policy_improvement(iii_V, P, gamma=0.99)\n",
    "print_policy(iiii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, iiii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, iiii_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00   0.52 | 01   0.38 | 02   0.26 | 03   0.25 |\n",
      "| 04   0.54 |           | 06   0.28 |           |\n",
      "| 08   0.57 | 09   0.62 | 10   0.58 |           |\n",
      "|           | 13   0.72 | 14   0.85 |           |\n"
     ]
    }
   ],
   "source": [
    "iiii_V = policy_evaluation(iiii_pi, P, gamma=0.99)\n",
    "print_state_value_function(iiii_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      < | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n"
     ]
    }
   ],
   "source": [
    "iiiii_pi = policy_improvement(iiii_V, P, gamma=0.99)\n",
    "print_policy(iiiii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, iiiii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, iiiii_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00   0.53 | 01   0.45 | 02   0.38 | 03   0.37 |\n",
      "| 04   0.55 |           | 06   0.32 |           |\n",
      "| 08   0.58 | 09   0.63 | 10    0.6 |           |\n",
      "|           | 13   0.73 | 14   0.86 |           |\n"
     ]
    }
   ],
   "source": [
    "iiiii_V = policy_evaluation(iiiii_pi, P, gamma=0.99)\n",
    "print_state_value_function(iiiii_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n"
     ]
    }
   ],
   "source": [
    "iiiiii_pi = policy_improvement(iiiii_V, P, gamma=0.99)\n",
    "print_policy(iiiiii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, iiiiii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, iiiiii_pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00   0.54 | 01    0.5 | 02   0.47 | 03   0.46 |\n",
      "| 04   0.56 |           | 06   0.36 |           |\n",
      "| 08   0.59 | 09   0.64 | 10   0.62 |           |\n",
      "|           | 13   0.74 | 14   0.86 |           |\n"
     ]
    }
   ],
   "source": [
    "iiiiii_V = policy_evaluation(iiiiii_pi, P, gamma=0.99)\n",
    "print_state_value_function(iiiiii_V, P, prec=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n"
     ]
    }
   ],
   "source": [
    "iiiiiii_pi = policy_improvement(iiiiii_V, P, gamma=0.99)\n",
    "print_policy(iiiiiii_pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, iiiiiii_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, iiiiiii_pi)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 정책 반복법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "상태-가치 함수:\n",
      "| 00  0.542 | 01 0.4988 | 02 0.4707 | 03 0.4569 |\n",
      "| 04 0.5585 |           | 06 0.3583 |           |\n",
      "| 08 0.5918 | 09 0.6431 | 10 0.6152 |           |\n",
      "|           | 13 0.7417 | 14 0.8628 |           |\n",
      "\n",
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n"
     ]
    }
   ],
   "source": [
    "V_best_p, pi_best_p = policy_iteration(P, gamma=0.99)\n",
    "print_state_value_function(V_best_p, P, prec=4)\n",
    "print()\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best_p, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best_p, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best_p)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Slippery Walk Five"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "env = gym.make('SlipperyWalkFive-v0')\n",
    "init_state = env.reset()\n",
    "goal_state = 6\n",
    "P = env.env.P"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 가치 반복법"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "def value_iteration(P, gamma=1.0, theta=1e-10):\n",
    "    V = np.zeros(len(P), dtype=np.float64)\n",
    "    while True:\n",
    "        Q = np.zeros((len(P), len(P[0])), dtype=np.float64)\n",
    "        for s in range(len(P)):\n",
    "            for a in range(len(P[s])):\n",
    "                for prob, next_state, reward, done in P[s][a]:\n",
    "                    Q[s][a] += prob * (reward + gamma * V[next_state] * (not done))\n",
    "        if np.max(np.abs(V - np.max(Q, axis=1))) < theta:\n",
    "            break\n",
    "        V = np.max(Q, axis=1)\n",
    "    pi = lambda s: {s:a for s, a in enumerate(np.argmax(Q, axis=1))}[s]\n",
    "    return V, pi"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "|           | 01      > | 02      > | 03      > | 04      > | 05      > |           |\n",
      "Reaches goal 93.00%. Obtains an average undiscounted return of 0.9300.\n",
      "\n",
      "상태-가치 함수:\n",
      "|           | 01 0.66758 | 02 0.89011 | 03 0.96429 | 04 0.98901 | 05 0.99725 |           |\n"
     ]
    }
   ],
   "source": [
    "optimal_V, optimal_pi = value_iteration(P)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(optimal_pi, P, action_symbols=('<', '>'), n_cols=7)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, optimal_pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, optimal_pi)))\n",
    "print()\n",
    "print_state_value_function(optimal_V, P, n_cols=7, prec=5)\n",
    "# |            | 01   0.668 | 02   0.890 | 03   0.964 | 04   0.989 | 05   0.997 |            |"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Frozen Lake MDP"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "env = gym.make('FrozenLake-v0')\n",
    "init_state = env.reset()\n",
    "goal_state = 15\n",
    "P = env.env.P"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (VI):\n",
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.542 | 01 0.4988 | 02 0.4707 | 03 0.4569 |\n",
      "| 04 0.5585 |           | 06 0.3583 |           |\n",
      "| 08 0.5918 | 09 0.6431 | 10 0.6152 |           |\n",
      "|           | 13 0.7417 | 14 0.8628 |           |\n"
     ]
    }
   ],
   "source": [
    "V_best_v, pi_best_v = value_iteration(P, gamma=0.99)\n",
    "print('Optimal policy and state-value function (VI):')\n",
    "print_policy(pi_best_v, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best_v, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best_v)))\n",
    "print()\n",
    "print_state_value_function(V_best_v, P, prec=4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "For comparison, optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      < | 01      ^ | 02      ^ | 03      ^ |\n",
      "| 04      < |           | 06      < |           |\n",
      "| 08      ^ | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      v |           |\n",
      "Reaches goal 74.00%. Obtains an average undiscounted return of 0.7400.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.542 | 01  0.499 | 02  0.471 | 03  0.457 |\n",
      "| 04  0.558 |           | 06  0.358 |           |\n",
      "| 08  0.592 | 09  0.643 | 10  0.615 |           |\n",
      "|           | 13  0.742 | 14  0.863 |           |\n"
     ]
    }
   ],
   "source": [
    "print('For comparison, optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best_p, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best_p, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best_p)))\n",
    "print()\n",
    "print_state_value_function(V_best_p, P)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Changing the Frozen Lake environment MDP"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "env = gym.make('FrozenLake-v0')\n",
    "P = env.env.P\n",
    "\n",
    "# change reward function\n",
    "reward_goal, reward_holes, reward_others = 1, -1, -0.01\n",
    "goal, hole = 15, [5, 7, 11, 12]\n",
    "for s in range(len(P)):\n",
    "    for a in range(len(P[s])):\n",
    "        for t in range(len(P[s][a])):\n",
    "            values = list(P[s][a][t])\n",
    "            if values[1] == goal:\n",
    "                values[2] = reward_goal\n",
    "                values[3] = False\n",
    "            elif values[1] in hole:\n",
    "                values[2] = reward_holes\n",
    "                values[3] = False\n",
    "            else:\n",
    "                values[2] = reward_others\n",
    "                values[3] = False\n",
    "            if s in hole or s == goal:\n",
    "                values[2] = 0\n",
    "                values[3] = True\n",
    "            P[s][a][t] = tuple(values)\n",
    "\n",
    "# change transition function\n",
    "prob_action, prob_drift_one, prob_drift_two = 0.8, 0.1, 0.1\n",
    "for s in range(len(P)):\n",
    "    for a in range(len(P[s])):\n",
    "        for t in range(len(P[s][a])):\n",
    "            if P[s][a][t][0] == 1.0:\n",
    "                continue\n",
    "            values = list(P[s][a][t])\n",
    "            if t == 0:\n",
    "                values[0] = prob_drift_one\n",
    "            elif t == 1:\n",
    "                values[0] = prob_action\n",
    "            elif t == 2:\n",
    "                values[0] = prob_drift_two\n",
    "            P[s][a][t] = tuple(values)\n",
    "\n",
    "env.env.P = P"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      v | 01      ^ | 02      v | 03      ^ |\n",
      "| 04      < |           | 06      v |           |\n",
      "| 08      > | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      > |           |\n",
      "Reaches goal 78.00%. Obtains an average undiscounted return of 0.3657.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.433 | 01  0.353 | 02  0.409 | 03   0.28 |\n",
      "| 04  0.461 |           | 06   0.45 |           |\n",
      "| 08  0.636 | 09  0.884 | 10  0.831 |           |\n",
      "|           | 13  0.945 | 14  0.977 |           |\n"
     ]
    }
   ],
   "source": [
    "V_best, pi_best = policy_iteration(env.env.P, gamma=0.99)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best)))\n",
    "print()\n",
    "print_state_value_function(V_best, P)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      v | 01      ^ | 02      v | 03      ^ |\n",
      "| 04      < |           | 06      v |           |\n",
      "| 08      > | 09      v | 10      < |           |\n",
      "|           | 13      > | 14      > |           |\n",
      "Reaches goal 78.00%. Obtains an average undiscounted return of 0.3657.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.433 | 01  0.353 | 02  0.409 | 03   0.28 |\n",
      "| 04  0.461 |           | 06   0.45 |           |\n",
      "| 08  0.636 | 09  0.884 | 10  0.831 |           |\n",
      "|           | 13  0.945 | 14  0.977 |           |\n"
     ]
    }
   ],
   "source": [
    "V_best, pi_best = value_iteration(env.env.P, gamma=0.99)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best)))\n",
    "print()\n",
    "print_state_value_function(V_best, P)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Russell & Norvig's Gridworld"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [],
   "source": [
    "env = gym.make('RussellNorvigGridworld-v0')\n",
    "init_state = env.reset()\n",
    "goal_state = 3\n",
    "P = env.env.P"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      > | 01      > | 02      > |           |\n",
      "| 04      ^ |           | 06      ^ |           |\n",
      "| 08      ^ | 09      < | 10      < | 11      < |\n",
      "Reaches goal 96.00%. Obtains an average undiscounted return of 0.6424.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.812 | 01  0.868 | 02  0.918 |           |\n",
      "| 04  0.762 |           | 06   0.66 |           |\n",
      "| 08  0.705 | 09  0.655 | 10  0.611 | 11  0.388 |\n"
     ]
    }
   ],
   "source": [
    "V_best_p, pi_best = policy_iteration(P)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best)))\n",
    "print()\n",
    "print_state_value_function(V_best_p, P)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Optimal policy and state-value function (PI):\n",
      "정책:\n",
      "| 00      > | 01      > | 02      > |           |\n",
      "| 04      ^ |           | 06      ^ |           |\n",
      "| 08      ^ | 09      < | 10      < | 11      < |\n",
      "Reaches goal 96.00%. Obtains an average undiscounted return of 0.6424.\n",
      "\n",
      "상태-가치 함수:\n",
      "| 00  0.812 | 01  0.868 | 02  0.918 |           |\n",
      "| 04  0.762 |           | 06   0.66 |           |\n",
      "| 08  0.705 | 09  0.655 | 10  0.611 | 11  0.388 |\n"
     ]
    }
   ],
   "source": [
    "V_best_v, pi_best = value_iteration(P)\n",
    "print('Optimal policy and state-value function (PI):')\n",
    "print_policy(pi_best, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi_best, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi_best)))\n",
    "print()\n",
    "print_state_value_function(V_best_v, P)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Re-construct optimal policy:\n",
      "정책:\n",
      "| 00      > | 01      > | 02      > |           |\n",
      "| 04      ^ |           | 06      ^ |           |\n",
      "| 08      ^ | 09      < | 10      < | 11      < |\n",
      "Reaches goal 96.00%. Obtains an average undiscounted return of 0.6424.\n"
     ]
    }
   ],
   "source": [
    "LEFT, DOWN, RIGHT, UP = range(4)\n",
    "pi = lambda s: {\n",
    "    0:RIGHT, 1:RIGHT, 2:RIGHT, 3:LEFT,\n",
    "    4:UP, 5:LEFT, 6:UP, 7:LEFT,\n",
    "    8:UP, 9:LEFT, 10:LEFT, 11:LEFT\n",
    "}[s]\n",
    "print('Re-construct optimal policy:')\n",
    "print_policy(pi, P)\n",
    "print('Reaches goal {:.2f}%. Obtains an average undiscounted return of {:.4f}.'.format(\n",
    "    probability_success(env, pi, goal_state=goal_state)*100, \n",
    "    mean_return(env, pi)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Evaluate optimal policy:\n",
      "상태-가치 함수:\n",
      "| 00  0.812 | 01  0.868 | 02  0.918 |           |\n",
      "| 04  0.762 |           | 06   0.66 |           |\n",
      "| 08  0.705 | 09  0.655 | 10  0.611 | 11  0.388 |\n"
     ]
    }
   ],
   "source": [
    "V = policy_evaluation(pi, P)\n",
    "print('Evaluate optimal policy:')\n",
    "print_state_value_function(V, P)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Improve optimal policy (nothing to improve -- it is the same policy, because it is optimal):\n",
      "정책:\n",
      "| 00      > | 01      > | 02      > |           |\n",
      "| 04      ^ |           | 06      ^ |           |\n",
      "| 08      ^ | 09      < | 10      < | 11      < |\n"
     ]
    }
   ],
   "source": [
    "pi = policy_improvement(V, P)\n",
    "print('Improve optimal policy (nothing to improve -- it is the same policy, because it is optimal):')\n",
    "print_policy(pi, P)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "There are no differences, nothing to improve on the optimal policy and state-value function:\n",
      "[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n",
      "[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n"
     ]
    }
   ],
   "source": [
    "print('There are no differences, nothing to improve on the optimal policy and state-value function:')\n",
    "print(np.abs(V_best_p - V))\n",
    "print(np.abs(V_best_v - V))"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [],
   "name": "PyTorch_Tutorial.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}