\uac15\ud654\ud559\uc2b5(Reinforcement Learning, RL)\uc740 \uba38\uc2e0\ub7ec\ub2dd\uc758 \ud55c \ubd84\uc57c\ub85c, \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba74\uc11c \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\ub860\uc785\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 Python\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac15\ud654\ud559\uc2b5 \ud658\uacbd\uc744 \uad6c\ucd95\ud558\ub294 \ubc29\ubc95\uacfc \uae30\ubcf8\uc801\uc778 RL \uc5d0\uc774\uc804\ud2b8\ub97c \uad6c\ud604\ud558\uace0 \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uad6c\uc131 \uc694\uc18c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\uac15\ud654\ud559\uc2b5\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c Python \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\ub824\uba74 \ub2e4\uc74c \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc2ed\uc2dc\uc624:<\/p>\n
pip install numpy gym matplotlib<\/code><\/pre>\n3. \uac15\ud654\ud559\uc2b5 \ud658\uacbd \uad6c\ucd95<\/h2>\n
\uc774\uc81c OpenAI Gym\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c \uac15\ud654\ud559\uc2b5 \ud658\uacbd\uc744 \uad6c\ucd95\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uac00\uc7a5 \uae30\ubcf8\uc801\uc778 \ud658\uacbd \uc911 \ud558\ub098\ub294 CartPole<\/strong>\uc785\ub2c8\ub2e4. \uc774 \ud658\uacbd\uc740 \ub9c9\ub300\uae30\uac00 \uc138\uc6cc\uc9c4 \uc218\uc9c1 \ub9c9\ub300\ub97c \uc911\uc2ec\uc73c\ub85c \uade0\ud615\uc744 \uc7a1\ub294 \uac83\uc774 \ubaa9\ud45c\uc785\ub2c8\ub2e4.<\/p>\n3.1 CartPole \ud658\uacbd \ub9cc\ub4e4\uae30<\/h3>\n\nimport gym\n\n# CartPole \ud658\uacbd \uc0dd\uc131\nenv = gym.make('CartPole-v1')\nobs = env.reset()\n\n# \ud658\uacbd\uc744 5\ubc88 \uc2e4\ud589\ud574\ubcf4\uae30\nfor _ in range(5):\n env.render() # \ud658\uacbd \uc2dc\uac01\ud654\n obs, reward, done, info = env.step(env.action_space.sample()) # \ubb34\uc791\uc704 \ud589\ub3d9 \uc120\ud0dd\n if done:\n obs = env.reset() # \ud658\uacbd \ucd08\uae30\ud654\nenv.close()\n <\/code><\/pre>\n\uc704 \ucf54\ub4dc\ub97c \uc2e4\ud589\ud558\uba74 CartPole \ud658\uacbd\uc774 \uc0dd\uc131\ub418\uace0, \uc5d0\uc774\uc804\ud2b8\uac00 \ub79c\ub364 \ud589\ub3d9\uc744 \ud1b5\ud574 \ud658\uacbd\uc744 \ud0d0\uc0c9\ud558\ub294 \ubaa8\uc2b5\uc744 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3.2 \uc0ac\uc6a9\uc790 \uc815\uc758 \ud658\uacbd \ub9cc\ub4e4\uae30<\/h3>\n
OpenAI Gym\uc740 \ub2e4\uc591\ud55c \uae30\ubcf8 \ud658\uacbd\uc744 \uc81c\uacf5\ud558\uc9c0\ub9cc, \ub54c\ub54c\ub85c \uac1c\uc778\uc758 \ud544\uc694\uc5d0 \ub9de\ub294 \uc0ac\uc6a9\uc790 \uc815\uc758 \ud658\uacbd\uc774 \ud544\uc694\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc0ac\uc6a9\uc790 \uc815\uc758 \ud658\uacbd\uc744 \ub9cc\ub4e4\uae30 \uc704\ud574\uc11c\ub294 gym.Env<\/strong> \ud074\ub798\uc2a4\ub97c \uc0c1\uc18d\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ud074\ub798\uc2a4\ub97c \uc815\uc758\ud558\uba74 \ub429\ub2c8\ub2e4.<\/p>\n\nimport gym\nfrom gym import spaces\n\nclass CustomEnv(gym.Env):\n def __init__(self):\n super(CustomEnv, self).__init__()\n self.action_space = spaces.Discrete(2) # 0 \ub610\ub294 1 \uc120\ud0dd\n self.observation_space = spaces.Box(low=0, high=1, shape=(1,), dtype=float)\n self.state = 0\n\n def step(self, action):\n if action == 1:\n self.state = 1\n reward = 1\n else:\n self.state = 0\n reward = 0\n return self.state, reward, False, {}\n\n def reset(self):\n self.state = 0\n return self.state\n\n# \ud658\uacbd \uc0ac\uc6a9 \uc608\nenv = CustomEnv()\nobs = env.reset()\nprint(obs)\n <\/code><\/pre>\n\uc704\uc758 \ucf54\ub4dc \uc608\uc81c\ub294 \uac04\ub2e8\ud55c \uc0ac\uc6a9\uc790 \uc815\uc758 \ud658\uacbd\uc744 \uc0dd\uc131\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc774 \ud658\uacbd\uc740 \ub450 \uac00\uc9c0 \ud589\ub3d9(0 \ub610\ub294 1)\uc744 \uc120\ud0dd\ud558\uace0, \uadf8\uc5d0 \ub530\ub77c \ubcf4\uc0c1\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
4. \uae30\ubcf8 RL \uc5d0\uc774\uc804\ud2b8 \uad6c\ud604<\/h2>\n
\uc774\uc81c \uc5d0\uc774\uc804\ud2b8\ub97c \uad6c\ud604\ud560 \ucc28\ub840\uc785\ub2c8\ub2e4. \uac00\uc7a5 \uae30\ubcf8\uc801\uc778 RL \uc54c\uace0\ub9ac\uc998\uc778 Q-Learning\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc5d0\uc774\uc804\ud2b8\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
4.1 Q-Learning \uc774\ud574\ud558\uae30<\/h3>\n
Q-Learning\uc740 \ubaa8\ub378\ud504\ub9ac \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc73c\ub85c, \uc5d0\uc774\uc804\ud2b8\ub294 \uc0c1\ud0dc-\ud589\ub3d9 \uc30d\uc5d0 \ub300\ud55c Q\uac12\uc744 \ud559\uc2b5\ud558\uc5ec \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4. Q\uac12\uc740 \uc0c1\ud0dc\uc640 \ud589\ub3d9\uc5d0 \ub300\ud55c \ubcf4\uc0c1\uc758 \uae30\ub300\uac12\uc744 \ub098\ud0c0\ub0c5\ub2c8\ub2e4.<\/p>\n
4.2 Q-Learning \uc54c\uace0\ub9ac\uc998 \uad6c\ud604<\/h3>\n
\ub2e4\uc74c\uc740 Q-Learning \uc54c\uace0\ub9ac\uc998\uc744 \uc0ac\uc6a9\ud558\uc5ec CartPole \ud658\uacbd\uc5d0\uc11c \uc5d0\uc774\uc804\ud2b8\ub97c \ud559\uc2b5\ud558\ub294 \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n
\nimport numpy as np\n\n# Q-Learning \uc54c\uace0\ub9ac\uc998 \uad6c\ud604\nclass QLearningAgent:\n def __init__(self, n_states, n_actions, alpha=0.1, gamma=0.99, epsilon=1.0, epsilon_decay=0.995):\n self.q_table = np.zeros((n_states, n_actions))\n self.alpha = alpha # Learning rate\n self.gamma = gamma # Discount factor\n self.epsilon = epsilon # Explore rate\n self.epsilon_decay = epsilon_decay # Epsilon decay factor\n \n def choose_action(self, state):\n if np.random.rand() < self.epsilon:\n return np.random.randint(0, self.q_table.shape[1]) # Random action\n return np.argmax(self.q_table[state]) # Exploit learned values\n\n def update(self, state, action, reward, next_state):\n best_next_action = np.argmax(self.q_table[next_state])\n td_target = reward + self.gamma * self.q_table[next_state][best_next_action]\n td_delta = td_target - self.q_table[state][action]\n self.q_table[state][action] += self.alpha * td_delta\n\n# \uc5d0\uc774\uc804\ud2b8 \ud559\uc2b5 \uacfc\uc815\nenv = gym.make('CartPole-v1')\nn_states = 10 # \ub2e8\uc21c\ud654\ub97c \uc704\ud574 \uc0c1\ud0dc \uc218\ub97c \ucd95\uc18c\nn_actions = env.action_space.n\nagent = QLearningAgent(n_states, n_actions)\n\nfor episode in range(1000):\n state = env.reset()\n done = False\n \n while not done:\n action = agent.choose_action(state)\n next_state, reward, done, _ = env.step(action)\n agent.update(state, action, reward, next_state)\n state = next_state\n\nenv.close()\n <\/code><\/pre>\n\uc704\uc758 Q-Learning \uc5d0\uc774\uc804\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec CartPole \ud658\uacbd\uc5d0\uc11c \ud559\uc2b5\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. \uc5d0\ud53c\uc18c\ub4dc \uc218\ub97c \uc870\uc808\ud558\uc5ec \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uc5d0 \uc801\uc751\ud558\ub294 \ubaa8\uc2b5\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
5. \ud559\uc2b5 \uacfc\uc815 \uc2dc\uac01\ud654<\/h2>\n
\uc5d0\uc774\uc804\ud2b8\uc758 \ud559\uc2b5 \uacfc\uc815\uc744 \uc2dc\uac01\ud654\ud558\ub294 \uac83\uc740 \ub9e4\uc6b0 \uc911\uc694\ud569\ub2c8\ub2e4. Matplotlib\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc5d0\uc774\uc804\ud2b8\uac00 \uac01 \uc5d0\ud53c\uc18c\ub4dc\uc5d0\uc11c \uc5bb\uc740 \ucd1d \ubcf4\uc0c1\uc744 \uadf8\ub798\ud504\ub85c \ub098\ud0c0\ub0b4\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport matplotlib.pyplot as plt\n\nrewards = []\nfor episode in range(1000):\n # \uae30\uc874 \ud559\uc2b5 \ucf54\ub4dc\uc640 \ub3d9\uc77c\n total_reward = 0\n state = env.reset()\n done = False\n \n while not done:\n action = agent.choose_action(state)\n next_state, reward, done, _ = env.step(action)\n agent.update(state, action, reward, next_state)\n state = next_state\n total_reward += reward\n \n rewards.append(total_reward)\n\nplt.plot(rewards)\nplt.xlabel('Episode')\nplt.ylabel('Total Reward')\nplt.title('Training Progress')\nplt.show()\n <\/code><\/pre>\n\uc704 \ucf54\ub4dc\ub294 \uc5d0\uc774\uc804\ud2b8\uc758 \ud559\uc2b5 \uc9c4\ud589 \uc0c1\ud669\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uc744 \uc798 \ud0d0\uc0c9\ud560\uc218\ub85d \ucd1d \ubcf4\uc0c1\uc740 \uc99d\uac00\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
6. \uc2ec\ud654 \ud559\uc2b5: DQN<\/h2>\n
\uae30\ubcf8 Q-Learning\uc740 \uac04\ub2e8\ud558\uc9c0\ub9cc, \ub354 \ubcf5\uc7a1\ud55c \ud658\uacbd\uc5d0\uc11c\ub294 Deep Q-Networks (DQN)\uc640 \uac19\uc740 \uc2ec\uce35 \uac15\ud654\ud559\uc2b5 \ubc29\ubc95\uc774 \ub354 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4. DQN\uc740 \ub525\ub7ec\ub2dd\uc744 \uc0ac\uc6a9\ud558\uc5ec Q\uac12\uc744 \uadfc\uc0ac\ud569\ub2c8\ub2e4. \uc774 \ubd80\ubd84\uc5d0 \ub300\ud574\uc11c\ub294 \ubcc4\ub3c4\uc758 \uac15\uc88c\ub85c \ub2e4\ub8f0 \uc608\uc815\uc785\ub2c8\ub2e4.<\/p>\n
7. \uacb0\ub860<\/h2>\n
\uc774 \uae00\uc5d0\uc11c\ub294 Python\uc744 \ud65c\uc6a9\ud558\uc5ec \uac15\ud654\ud559\uc2b5 \ud658\uacbd\uc744 \uad6c\ucd95\ud558\uace0, Q-Learning \uc54c\uace0\ub9ac\uc998\uc744 \uc774\uc6a9\ud55c \uae30\ubcf8 \uc5d0\uc774\uc804\ud2b8\ub97c \uad6c\ud604\ud558\uba70 \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uac15\ud654\ud559\uc2b5\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0 \uc751\uc6a9\ub420 \uc218 \uc788\uc73c\uba70, \ubcf4\ub2e4 \ubc1c\uc804\ub41c \uc54c\uace0\ub9ac\uc998\uacfc \uae30\ubc95\ub4e4\ub3c4 \uc874\uc7ac\ud558\ubbc0\ub85c \uc55e\uc73c\ub85c\uc758 \ud559\uc2b5\uacfc \ud0d0\uc0c9\uc774 \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\uac15\ud654\ud559\uc2b5\uc740 \ub9e4\ub825\uc801\uc778 \ubd84\uc57c\ub85c, \ub2e4\uc591\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud55c \ub9ce\uc740 \uac00\ub2a5\uc131\uc744 \uc81c\uacf5\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\ub7ec\ubd84\uc758 \uc5f0\uad6c\uc640 \ud504\ub85c\uc81d\ud2b8\uc5d0 \ub3c4\uc6c0\uc774 \ub418\uae38 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n