\ub525\ub7ec\ub2dd\uc744 \ud65c\uc6a9\ud55c \ud050-\ub7ec\ub2dd\uc5d0 \ub300\ud55c \uae4a\uc774 \uc788\ub294 \ud0d0\uad6c<\/p>\n<\/header>\n \ud050-\ub7ec\ub2dd(QLearning)\uc740 \uac15\ud654 \ud559\uc2b5\uc758 \ud55c \ud615\ud0dc\ub85c, \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uc5ec \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \ud559\uc2b5\ud558\ub3c4\ub85d \ub3d5\uc2b5\ub2c8\ub2e4. \ud050-\ub7ec\ub2dd\uc758 \ud575\uc2ec \uc544\uc774\ub514\uc5b4\ub294 \uac01 \uc0c1\ud0dc(state)\uc5d0\uc11c \uac00\ub2a5\ud55c \ud589\ub3d9(action)\uc5d0 \ub300\ud55c \uac00\uce58(value)\ub97c \uc800\uc7a5\ud558\ub294 \ud050\ub97c \uc0ac\uc6a9\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub294 \uc5d0\uc774\uc804\ud2b8\uac00 \uc120\ud0dd\ud560 \uc218 \uc788\ub294 \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \uacb0\uc815\ud558\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4.<\/p>\n \ud050-\ub7ec\ub2dd\uc740 \uc77c\ubc18\uc801\uc73c\ub85c Markov Decision Process(MDP)\uc5d0 \uae30\ubc18\ud558\uba70, \ub2e4\uc74c\uacfc \uac19\uc740 \uc694\uc18c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n \ud050-\ub7ec\ub2dd \uc54c\uace0\ub9ac\uc998\uc740 Q\ud568\uc218\ub97c \uc5c5\ub370\uc774\ud2b8\ud558\ub294 \uae30\ubcf8 \uc544\uc774\ub514\uc5b4\ub97c \ud3ec\ud568\ud569\ub2c8\ub2e4. \uc5d0\uc774\uc804\ud2b8\ub294 \uac01 \uc2dc\uac04 \ub2e8\uacc4\uc5d0\uc11c \ub2e4\uc74c\uc758 \uc808\ucc28\ub97c \ub530\ub985\ub2c8\ub2e4:<\/p>\n Q\ud568\uc218 \uc5c5\ub370\uc774\ud2b8\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc218\uc2dd\uc73c\ub85c \ud45c\ud604\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n \uc5ec\uae30\uc11c \u03b1\ub294 \ud559\uc2b5\ub960, \u03b3\ub294 \ud560\uc778\uc728\uc744 \ub098\ud0c0\ub0c5\ub2c8\ub2e4. \uc774 \ub450 \uc694\uc18c\ub294 \uc5d0\uc774\uc804\ud2b8\uac00 \uacfc\uac70 \uacbd\ud5d8\uc744 \uc5bc\ub9c8\ub098 \ubc18\uc601\ud560\uc9c0\ub97c \uacb0\uc815\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n \uc774\uc81c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud050-\ub7ec\ub2dd\uc744 \uac04\ub2e8\ud558\uac8c \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 OpenAI\uc758 Gym \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud658\uacbd\uc744 \ub9cc\ub4e4\uace0, \ud050-\ub7ec\ub2dd \uc5d0\uc774\uc804\ud2b8\ub97c \ud559\uc2b5\uc2dc\ud0ac \uac83\uc785\ub2c8\ub2e4.<\/p>\n \ud050-\ub7ec\ub2dd\uc758 \uc8fc\uc694 \uc7a5\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n \ud558\uc9c0\ub9cc \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uc810\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n1. \ud050-\ub7ec\ub2dd\uc774\ub780?<\/h2>\n
\n
2. \ud050-\ub7ec\ub2dd \uc54c\uace0\ub9ac\uc998<\/h2>\n
\n
Q(S, A) <- Q(S, A) + \u03b1(R + \u03b3 * max(Q(S', A')) - Q(S, A))<\/code><\/pre>\n3. PyTorch\ub85c \ud050-\ub7ec\ub2dd \uad6c\ud604\ud558\uae30<\/h2>\n
import gym\nimport numpy as np\nimport random\n\n# \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130\nLEARNING_RATE = 0.1\nDISCOUNT_FACTOR = 0.9\nEPISODES = 1000\n\n# \ud658\uacbd \uc124\uc815\nenv = gym.make('Taxi-v3')\nQ_table = np.zeros([env.observation_space.n, env.action_space.n])\n\ndef select_action(state, epsilon):\n if random.uniform(0, 1) < epsilon:\n return env.action_space.sample() # \ub79c\ub364 \ud589\ub3d9 \uc120\ud0dd\n else:\n return np.argmax(Q_table[state]) # \uac00\uc7a5 \ub192\uc740 Q\uac12\uc758 \ud589\ub3d9 \uc120\ud0dd\n\nfor episode in range(EPISODES):\n state = env.reset()\n done = False\n epsilon = 1.0 \/ (episode \/ 100 + 1) # \ud0d0\uc0ac\uc728\n\n while not done:\n action = select_action(state, epsilon)\n next_state, reward, done, _ = env.step(action)\n \n # Q\ud568\uc218 \uc5c5\ub370\uc774\ud2b8\n Q_table[state][action] += LEARNING_RATE * (reward + DISCOUNT_FACTOR * np.max(Q_table[next_state]) - Q_table[state][action])\n \n state = next_state\n\nprint(\"\ud559\uc2b5 \uc644\ub8cc\")\n\n# \uc0d8\ud50c \ud14c\uc2a4\ud2b8\nstate = env.reset()\ndone = False\nwhile not done:\n action = np.argmax(Q_table[state]) # \ucd5c\uc801 \ud589\ub3d9 \uc120\ud0dd\n state, reward, done, _ = env.step(action)\n env.render() # \ud658\uacbd \ucd9c\ub825<\/code><\/pre>\n<\/section>\n4. \ud050-\ub7ec\ub2dd\uc758 \uc7a5\ub2e8\uc810<\/h2>\n
\n
\n