\ucd5c\uadfc \uba87 \ub144\uac04 \uc778\uacf5\uc9c0\ub2a5(AI) \uae30\uc220\uc758 \ubc1c\uc804\uc740 \ub180\ub77c\uc6b4 \uc18d\ub3c4\ub85c \uc774\ub8e8\uc5b4\uc84c\uace0, \uadf8\uc911\uc5d0\uc11c\ub3c4 \uac15\ud654\ud559\uc2b5\uc740 \ub9ce\uc740 \uc8fc\ubaa9\uc744 \ubc1b\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5(Policy-Based Reinforcement Learning) \uc54c\uace0\ub9ac\uc998\uc740 \uadf8 \ud6a8\uc728\uc131 \ubc0f \uc801\uc6a9 \uac00\ub2a5\uc131\uc73c\ub85c \uc778\ud574 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 PPO(Proximal Policy Optimization), A3C(Asynchronous Actor-Critic) \ubc0f SAC(Soft Actor-Critic) \uc54c\uace0\ub9ac\uc998\uc758 \uac1c\ub150\uc744 \uc0b4\ud3b4\ubcf4\uace0, \uc815\ucc45 \uae30\ubc18 \uc54c\uace0\ub9ac\uc998\uc758 \uc7a5\uc810\uc5d0 \ub300\ud574 \uc2ec\ub3c4 \uc788\uac8c \ub17c\uc758\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
1. \uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uac1c\ub150<\/h2>\n
\uac15\ud654\ud559\uc2b5\uc774\ub780, \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba74\uc11c \ubcf4\uc0c1\uc744 \ud1b5\ud574 \ucd5c\uc801\uc758 \ud589\ub3d9 \uc815\ucc45\uc744 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\ub860\uc785\ub2c8\ub2e4. \uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uc694\uc18c\ub85c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uac83\ub4e4\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
- \n
- \ud658\uacbd (Environment):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \uc0c1\ud638\uc791\uc6a9\ud558\ub294 \uacf3\uc785\ub2c8\ub2e4.<\/li>\n
- \uc5d0\uc774\uc804\ud2b8 (Agent):<\/strong> \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba70 \ud559\uc2b5\ud558\ub294 \uc8fc\uccb4\uc785\ub2c8\ub2e4.<\/li>\n
- \uc0c1\ud0dc (State):<\/strong> \ud658\uacbd\uc758 \ud604\uc7ac \uc0c1\ud669\uc744 \ub098\ud0c0\ub0c5\ub2c8\ub2e4.<\/li>\n
- \ud589\ub3d9 (Action):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \uc120\ud0dd\ud560 \uc218 \uc788\ub294 \ud589\ub3d9\uc785\ub2c8\ub2e4.<\/li>\n
- \ubcf4\uc0c1 (Reward):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \ud2b9\uc815 \ud589\ub3d9\uc744 \ud588\uc744 \ub54c \ubc1b\ub294 \ud53c\ub4dc\ubc31\uc785\ub2c8\ub2e4.<\/li>\n<\/ul>\n
\uac15\ud654\ud559\uc2b5\uc758 \ubaa9\ud45c\ub294 \uc5d0\uc774\uc804\ud2b8\uac00 \ud2b9\uc815 \uc0c1\ud0dc\uc5d0\uc11c \ucd5c\ub300\uc758 \ub204\uc801 \ubcf4\uc0c1\uc744 \ubc1b\uc744 \uc218 \uc788\ub3c4\ub85d \ucd5c\uc801\uc758 \uc815\ucc45\uc744 \ud559\uc2b5\ud558\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
2. \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5(PG)<\/h2>\n
\uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5\uc5d0\uc11c\ub294 \uc5d0\uc774\uc804\ud2b8\uc758 \ud589\ub3d9\uc744 \uc9c1\uc811\uc801\uc73c\ub85c \uc815\ud558\ub294 \uc815\ucc45\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc815\ucc45\uc774\ub780, \uc8fc\uc5b4\uc9c4 \uc0c1\ud0dc\uc5d0 \ub300\ud574 \uc5b4\ub5a4 \ud589\ub3d9\uc744 \ucde8\ud574\uc57c \ud558\ub294\uc9c0\ub97c \uc815\uc758\ud558\ub294 \ud568\uc218\uc785\ub2c8\ub2e4. \uc815\ucc45 \uae30\ubc18 \ud559\uc2b5\uc758 \uc7a5\uc810\uc740 \uc5f0\uc18d\uc801\uc778 \ub3d9\uc791 \uacf5\uac04\uc744 \uac00\uc9c4 \ubb38\uc81c\uc5d0\uc11c\ub3c4 \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub2e4\ub294 \uc810\uc785\ub2c8\ub2e4.<\/p>\n
3. PPO (Proximal Policy Optimization)<\/h2>\n
PPO\ub294 \ucd5c\uc2e0\uc758 \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998 \uc911 \ud558\ub098\ub85c, \uae30\uc874\uc758 \uc815\ucc45 \uc5c5\ub370\uc774\ud2b8 \ubc29\ubc95\uc5d0 \ube44\ud574 \ub354 \uc548\uc815\uc801\uc778 \ud559\uc2b5 \uacfc\uc815\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. PPO\ub294 \ud074\ub9ac\ud551(clipping) \uae30\uc220\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc815\ucc45 \ubcc0\ud654\uc758 \ud06c\uae30\ub97c \uc81c\ud55c\ud569\ub2c8\ub2e4. \uc774\ub294 \uc815\ucc45\uc774 \uae09\uaca9\ud788 \ubcc0\ud654\ud558\ub294 \uac83\uc744 \ubc29\uc9c0\ud558\uba70, \uc548\uc815\uc801\uc778 \ud559\uc2b5\uc744 \uc720\ub3c4\ud569\ub2c8\ub2e4.<\/p>\n
3.1 PPO\uc758 \uc8fc\uc694 \uac1c\ub150<\/h3>\n
PPO\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc8fc\uc694 \uac1c\ub150\uc744 \ud3ec\ud568\ud569\ub2c8\ub2e4:<\/p>\n
- \n
- \ud074\ub9ac\ud551 (Clipping):<\/strong> \uc815\ucc45\uc774 \ubcc0\ud654\ud558\ub294 \ubc94\uc704\ub97c \uc81c\ud55c\ud558\uc5ec \uc548\uc815\uc131\uc744 \ub192\uc785\ub2c8\ub2e4.<\/li>\n
- \uae30\ub300 \ubcf4\uc0c1 (Expected Reward):<\/strong> \ub2e4\uc591\ud55c \ud589\ub3d9\uc5d0 \ub300\ud574 \uae30\ub300\ud560 \uc218 \uc788\ub294 \ubcf4\uc0c1\uc744 \uae30\ubc18\uc73c\ub85c \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/li>\n
- \uc2e0\ub8b0 \uad6c\uac04 (Trust Region):<\/strong> \ud3f4\ub9ac\uc2dc\uac00 \uae30\uc874 \uc815\ucc45\uacfc \ub108\ubb34 \uba40\ub9ac \ubc97\uc5b4\ub098\uc9c0 \uc54a\ub3c4\ub85d \ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
3.2 PPO \uc608\uc81c \uc18c\uc2a4 \ucf54\ub4dc<\/h3>\n
\nimport gym\nimport numpy as np\nimport tensorflow as tf\n\nclass PPO:\n def __init__(self, state_size, action_size):\n self.state_size = state_size\n self.action_size = action_size\n self.learning_rate = 0.001\n self.gamma = 0.99 # discount factor\n \n # Build policy model\n self.model = self.build_model()\n \n def build_model(self):\n model = tf.keras.Sequential([\n tf.keras.layers.Dense(24, activation='relu', input_shape=(self.state_size,)),\n tf.keras.layers.Dense(24, activation='relu'),\n tf.keras.layers.Dense(self.action_size, activation='softmax')\n ])\n model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=self.learning_rate), loss='categorical_crossentropy')\n return model\n\n def choose_action(self, state):\n state = state.reshape([1, self.state_size])\n probabilities = self.model.predict(state).flatten()\n action = np.random.choice(self.action_size, p=probabilities)\n return action\n\n def train(self, states, actions, rewards):\n # Placeholder for training logic\n pass\n\n# \ud658\uacbd \ucd08\uae30\ud654\nenv = gym.make('CartPole-v1')\nstate_size = env.observation_space.shape[0]\naction_size = env.action_space.n\nppo_agent = PPO(state_size, action_size)\n\n# \uc5d0\ud53c\uc18c\ub4dc \ub8e8\ud504\nfor episode in range(1000):\n state = env.reset()\n done = False\n while not done:\n action = ppo_agent.choose_action(state)\n next_state, reward, done, _ = env.step(action)\n ppo_agent.train(state, action, reward)\n state = next_state\n if done:\n break\n <\/code><\/pre>\n4. A3C (Asynchronous Actor-Critic)<\/h2>\n
A3C\ub294 \ube44\ub3d9\uae30\uc801 \ubc29\ubc95\ub860\uc744 \uc0ac\uc6a9\ud558\ub294 \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc73c\ub85c, \ub2e4\uc218\uc758 \uc5d0\uc774\uc804\ud2b8\ub97c \ubcd1\ub82c\ub85c \uc2e4\ud589\ud558\uc5ec \ud559\uc2b5 \ud6a8\uc728\uc744 \ub192\uc785\ub2c8\ub2e4. \uac01 \uc5d0\uc774\uc804\ud2b8\ub294 \ubcc4\ub3c4\uc758 \ud658\uacbd\uc5d0\uc11c \ub3c5\ub9bd\uc801\uc73c\ub85c \uc791\ub3d9\ud558\uba70, \uc5ec\ub7ec \uc5d0\uc774\uc804\ud2b8\uc758 \uacbd\ud5d8\uc744 \ud1b5\ud569\ud558\uc5ec \uc5c5\ub370\uc774\ud2b8\ub97c \uc218\ud589\ud569\ub2c8\ub2e4.<\/p>\n
4.1 A3C\uc758 \uc8fc\uc694 \uac1c\ub150<\/h3>\n
- \n
- \ube44\ub3d9\uae30 \uc2e4\ud589 (Asynchronous Execution):<\/strong> \uc5ec\ub7ec \uc778\uc2a4\ud134\uc2a4\uac00 \ub3d9\uc2dc\uc5d0 \uc2e4\ud589\ub418\uc5b4 \ub370\uc774\ud130\uac00 \ub354 \ud48d\ubd80\ud574\uc9d1\ub2c8\ub2e4.<\/li>\n
- \uc561\ud130-\ube44\ud3c9\uac00 \uc544\ud0a4\ud14d\ucc98 (Actor-Critic Architecture):<\/strong> \uc5d0\uc774\uc804\ud2b8\ub294 \ub450 \uac00\uc9c0 \uc5ed\ud560\uc744 \ub9e1\uc544 \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uace0, \uadf8 \ud589\ub3d9\uc758 \uac00\uce58\ub97c \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/li>\n
- \uacbd\ud5d8 \uc7ac\uc0ac\uc6a9 (Experience Replay):<\/strong> \uc774\uc804 \uacbd\ud5d8\uc744 \uba54\ubaa8\ub9ac\uc5d0 \uc800\uc7a5\ud558\uace0 \uc5ec\ub7ec \ubc88 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
4.2 A3C \uc608\uc81c \uc18c\uc2a4 \ucf54\ub4dc<\/h3>\n
\nimport threading\nimport gym\nimport numpy as np\nimport tensorflow as tf\n\nclass A3C:\n def __init__(self, state_size, action_size):\n self.state_size = state_size\n self.action_size = action_size\n self.model = self.build_model()\n \n def build_model(self):\n model = tf.keras.Sequential([\n tf.keras.layers.Dense(24, activation='relu', input_shape=(self.state_size,)),\n tf.keras.layers.Dense(24, activation='relu'),\n tf.keras.layers.Dense(self.action_size, activation='softmax')\n ])\n return model\n \n def choose_action(self, state):\n state = state.reshape([1, self.state_size])\n probabilities = self.model.predict(state).flatten()\n action = np.random.choice(self.action_size, p=probabilities)\n return action\n\n def train(self, states, actions, rewards):\n # Placeholder for training logic\n pass\n\ndef worker(env_name, global_model, state_size, action_size):\n env = gym.make(env_name)\n local_model = A3C(state_size, action_size)\n\n while True:\n state = env.reset()\n done = False\n while not done:\n action = local_model.choose_action(state)\n next_state, reward, done, _ = env.step(action)\n local_model.train(state, action, reward)\n state = next_state\n if done:\n break\n\n# \ud658\uacbd \uc124\uc815\nenv_name = 'CartPole-v1'\nstate_size = 4\naction_size = 2\nglobal_model = A3C(state_size, action_size)\n\n# \uc2a4\ub808\ub4dc\ub85c A3C \uc5d0\uc774\uc804\ud2b8 \uc2e4\ud589\nthreads = []\nfor i in range(4):\n t = threading.Thread(target=worker, args=(env_name, global_model, state_size, action_size))\n threads.append(t)\n t.start()\n\nfor t in threads:\n t.join()\n <\/code><\/pre>\n5. SAC (Soft Actor-Critic)<\/h2>\n
SAC\ub294 \uc624\ud504\ub77c\uc778 \ud559\uc2b5 \ubc29\uc2dd\uc744 \uc0ac\uc6a9\ud558\ub294 \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc73c\ub85c, \ucd5c\ub300 \uc5d4\ud2b8\ub85c\ud53c \uc6d0\uce59\uc744 \ud65c\uc6a9\ud558\uc5ec \uc5d0\uc774\uc804\ud2b8\uac00 \ub354 \ud0d0\ud5d8\ud558\ub3c4\ub85d \uc720\ub3c4\ud569\ub2c8\ub2e4. \uc774 \uc54c\uace0\ub9ac\uc998\uc740 \uc8fc\ub85c \uc5f0\uc18d\uc801\uc778 \ud589\ub3d9 \uacf5\uac04\uc744 \uac00\uc9c4 \ubb38\uc81c\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec \uc90d\ub2c8\ub2e4.<\/p>\n
5.1 SAC\uc758 \uc8fc\uc694 \uac1c\ub150<\/h3>\n
- \n
- \ucd5c\ub300 \uc5d4\ud2b8\ub85c\ud53c (Maximum Entropy):<\/strong> \ud589\ub3d9\uc758 \ubd88\ud655\uc2e4\uc131\uc744 \uadf9\ub300\ud654\ud558\uc5ec \ud0d0\ud5d8\uc744 \uc7a5\ub824\ud569\ub2c8\ub2e4.<\/li>\n
- \uc624\ud504\ub77c\uc778 \ud559\uc2b5 (Off-Policy Learning):<\/strong> \uc774\uc804 \uacbd\ud5d8\uc744 \uc7ac\uc0ac\uc6a9\ud558\uc5ec \ud6a8\uc728\uc801\uc73c\ub85c \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/li>\n
- \ube44\ud310\uc790 (Critic):<\/strong> \uac00\uce58 \ud3c9\uac00\ub97c \uc218\ud589\ud558\uc5ec \uc815\ucc45 \ud559\uc2b5\uc744 \uc9c0\uc6d0\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
5.2 SAC \uc608\uc81c \uc18c\uc2a4 \ucf54\ub4dc<\/h3>\n
\nimport gym\nimport numpy as np\nimport tensorflow as tf\n\nclass SAC:\n def __init__(self, state_size, action_size):\n self.state_size = state_size\n self.action_size = action_size\n self.learning_rate = 0.001\n \n # \uac01\uc885 \ubaa8\ub378 \uc815\uc758\n self.policy_model = self.build_policy_model()\n self.value_model = self.build_value_model()\n \n def build_policy_model(self):\n model = tf.keras.Sequential([\n tf.keras.layers.Dense(256, activation='relu', input_shape=(self.state_size,)),\n tf.keras.layers.Dense(self.action_size, activation='tanh')\n ])\n model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=self.learning_rate))\n return model\n \n def build_value_model(self):\n model = tf.keras.Sequential([\n tf.keras.layers.Dense(256, activation='relu', input_shape=(self.state_size,)),\n tf.keras.layers.Dense(1)\n ])\n model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=self.learning_rate))\n return model\n\n def choose_action(self, state):\n state = state.reshape([1, self.state_size])\n action = self.policy_model.predict(state)\n return action.flatten()\n\n def train(self, states, actions, rewards):\n # Placeholder for training logic\n pass\n\n# \ud658\uacbd \ucd08\uae30\ud654\nenv = gym.make('Pendulum-v0')\nstate_size = env.observation_space.shape[0]\naction_size = env.action_space.shape[0]\nsac_agent = SAC(state_size, action_size)\n\n# \uc5d0\ud53c\uc18c\ub4dc \ub8e8\ud504\nfor episode in range(1000):\n state = env.reset()\n done = False\n while not done:\n action = sac_agent.choose_action(state)\n next_state, reward, done, _ = env.step(action)\n sac_agent.train(state, action, reward)\n state = next_state\n if done:\n break\n <\/code><\/pre>\n6. \uc815\ucc45 \uae30\ubc18 \uc54c\uace0\ub9ac\uc998\uc758 \uc7a5\uc810<\/h2>\n
\uc815\ucc45 \uae30\ubc18 \uc54c\uace0\ub9ac\uc998\uc740 \uc5ec\ub7ec \uac00\uc9c0 \uc7a5\uc810\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
- \n
- \uc5f0\uc18d\uc801\uc778 \ud589\ub3d9 \uacf5\uac04:<\/strong> \uc815\ucc45 \uae30\ubc18 \ubc29\ubc95\uc740 \uc5f0\uc18d\uc801\uc778 \ud589\ub3d9 \uacf5\uac04\uc5d0\uc11c \ub9e4\uc6b0 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/li>\n
- \ud0d0\ud5d8\uacfc \ud65c\uc6a9:<\/strong> \uc815\ucc45 \uae30\ubc18 \uc54c\uace0\ub9ac\uc998\uc740 \uc5d0\uc774\uc804\ud2b8\uc758 \ud0d0\ud5d8\uc744 \uc720\ub3c4\ud558\ub294 \ub370 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/li>\n
- \uc9c1\uad00\uc801\uc778 \uc815\ucc45 \ud45c\ud604:<\/strong> \uc9c1\uc811\uc801\uc778 \uc815\ucc45 \ud559\uc2b5\uc73c\ub85c \uc778\ud574 \ubb38\uc81c\ub97c \uc774\ud574\ud558\uace0 \ud574\uacb0\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub429\ub2c8\ub2e4.<\/li>\n<\/ul>\n
\uacb0\ub860<\/h2>\n
PPO, A3C \ubc0f SAC\ub294 \ud604\uc7ac \uac00\uc7a5 \ud6a8\uacfc\uc801\uc778 \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998\uc73c\ub85c \ubd80\uac01\ub418\uace0 \uc788\uc73c\uba70, \uac01\uae30 \ub2e4\ub978 \uac15\uc810\uacfc \ud2b9\uc9d5\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub4e4 \uc54c\uace0\ub9ac\uc998\uc744 \ud1b5\ud574 \ub2e4\uc591\ud55c \ud658\uacbd\uc5d0\uc11c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uace0 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc815\ucc45 \uae30\ubc18 \uac15\ud654\ud559\uc2b5\uc758 \ubbf8\ub798\ub294 \ub354\uc6b1 \ubc1d\uc73c\uba70, \uc55e\uc73c\ub85c\ub3c4 \ub9ce\uc740 \uc5f0\uad6c\uc640 \ubc1c\uc804\uc774 \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n<\/article>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"
- \ud0d0\ud5d8\uacfc \ud65c\uc6a9:<\/strong> \uc815\ucc45 \uae30\ubc18 \uc54c\uace0\ub9ac\uc998\uc740 \uc5d0\uc774\uc804\ud2b8\uc758 \ud0d0\ud5d8\uc744 \uc720\ub3c4\ud558\ub294 \ub370 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/li>\n
- \uc624\ud504\ub77c\uc778 \ud559\uc2b5 (Off-Policy Learning):<\/strong> \uc774\uc804 \uacbd\ud5d8\uc744 \uc7ac\uc0ac\uc6a9\ud558\uc5ec \ud6a8\uc728\uc801\uc73c\ub85c \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/li>\n
- \uc561\ud130-\ube44\ud3c9\uac00 \uc544\ud0a4\ud14d\ucc98 (Actor-Critic Architecture):<\/strong> \uc5d0\uc774\uc804\ud2b8\ub294 \ub450 \uac00\uc9c0 \uc5ed\ud560\uc744 \ub9e1\uc544 \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uace0, \uadf8 \ud589\ub3d9\uc758 \uac00\uce58\ub97c \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/li>\n
- \uae30\ub300 \ubcf4\uc0c1 (Expected Reward):<\/strong> \ub2e4\uc591\ud55c \ud589\ub3d9\uc5d0 \ub300\ud574 \uae30\ub300\ud560 \uc218 \uc788\ub294 \ubcf4\uc0c1\uc744 \uae30\ubc18\uc73c\ub85c \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/li>\n
- \uc5d0\uc774\uc804\ud2b8 (Agent):<\/strong> \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba70 \ud559\uc2b5\ud558\ub294 \uc8fc\uccb4\uc785\ub2c8\ub2e4.<\/li>\n