\uac15\ud654\ud559\uc2b5\uc758 \ud575\uc2ec\uc740 \uc8fc\uc5b4\uc9c4 \ud658\uacbd\uc5d0\uc11c \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uc5ec \ucd5c\ub300\uc758 \ubcf4\uc0c1\uc744 \uc5bb\ub294 \ub3d9\uc548, \uc0c8\ub85c\uc6b4 \uc804\ub7b5\uc774\ub098 \uc0c1\ud0dc\ub97c \ud0d0\ud5d8\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud0d0\ud5d8(exploration)\uacfc \ud65c\uc6a9(exploitation)\uc758 \uade0\ud615\uc744 \ub9de\ucd94\ub294 \uac83\uc774 \ub9e4\uc6b0 \uc911\uc694\ud569\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 \u03b5-\uadf8\ub9ac\ub514(\u03b5-greedy), \uc18c\ud504\ud2b8\ub9e5\uc2a4(softmax) \ud0d0\ud5d8 \ubc29\ubc95\uc744 \ud3ec\ud568\ud55c \ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc18c\uac1c\ud558\uace0, \uc774\ub4e4\uc744 \ube44\uad50 \ubc0f \uc801\uc6a9\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc2ec\ub3c4 \uae4a\uc740 \ub17c\uc758\ub97c \ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud0d0\ud5d8\uacfc \ud65c\uc6a9\uc740 \uac15\ud654\ud559\uc2b5\uc5d0\uc11c \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\ub294 \ub450 \uac00\uc9c0 \uc8fc\uc694 \uc804\ub7b5\uc785\ub2c8\ub2e4. <\/p>\n
\uc774 \ub450 \uac00\uc9c0 \uc804\ub7b5\uc740 \uc801\uc808\ud558\uac8c \uc870\ud654\ub97c \uc774\ub8e8\uc5b4\uc57c\ub9cc \ucd5c\uc801\uc758 \uc131\uacfc\ub97c \ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc9c0\ub098\uce58\uac8c \ud0d0\ud5d8\uc744 \ud558\uac8c \ub418\uba74 \ubcf4\uc0c1\uc744 \uc5bb\uc9c0 \ubabb\ud560 \uc218 \uc788\uc73c\uba70, \ubc18\ub300\ub85c \ud65c\uc6a9\ub9cc \ud558\uac8c \ub418\uba74 \ucd5c\uc801\uc758 \ud589\ub3d9\uc774 \uc544\ub2cc \uc81c\ud55c\ub41c \uc815\ubcf4\ub85c \uc778\ud55c \ud558\uc704 \ucd5c\uc801\uc758 \ud574\uacb0\ucc45\uc5d0 \uba38\ubb3c\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
\uac15\ud654\ud559\uc2b5\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc5d0\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uac83\ub4e4\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\u03b5-\uadf8\ub9ac\ub514 \ubc29\ubc95\uc740 \uac00\uc7a5 \uae30\ucd08\uc801\uc774\uace0 \ub110\ub9ac \uc0ac\uc6a9\ub418\ub294 \ud0d0\ud5d8 \uc804\ub7b5 \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. \uc774 \uc804\ub7b5\uc5d0\uc11c \uc5d0\uc774\uc804\ud2b8\ub294 \u03b5\uc758 \ud655\ub960\ub85c \uc784\uc758\uc758 \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uace0, (1-\u03b5)\uc758 \ud655\ub960\ub85c \ud604\uc7ac \uac00\uc7a5 \uc88b\uc740 \ud589\ub3d9\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4.<\/p>\n
\uc608\ub97c \ub4e4\uc5b4, \u03b5 = 0.1\uc774\ub77c\uba74 \uc5d0\uc774\uc804\ud2b8\ub294 10%\uc758 \ud655\ub960\ub85c \ubb34\uc791\uc704 \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uace0, 90%\uc758 \ud655\ub960\ub85c \ud604\uc7ac \ucd94\uc815\ub41c \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc740 \ud0d0\ud5d8\uc758 \ube44\uc728\uc744 \uc870\uc808\ud560 \uc218 \uc788\ub294 \uc720\uc6a9\ud55c \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n
import numpy as np\n\nclass EpsilonGreedy:\n def __init__(self, n_actions, epsilon=0.1):\n self.n_actions = n_actions\n self.epsilon = epsilon\n self.action_values = np.zeros(n_actions)\n self.action_counts = np.zeros(n_actions)\n\n def select_action(self):\n if np.random.rand() < self.epsilon:\n return np.random.randint(self.n_actions) # \ud0d0\ud5d8\n else:\n return np.argmax(self.action_values) # \ud65c\uc6a9\n\n def update(self, action, reward):\n self.action_counts[action] += 1\n self.action_values[action] += (reward - self.action_values[action]) \/ self.action_counts[action]\n<\/code><\/pre>\n2.2 \uc18c\ud504\ud2b8\ub9e5\uc2a4 \ud0d0\ud5d8 (Softmax Exploration)<\/h3>\n
\uc18c\ud504\ud2b8\ub9e5\uc2a4 \ud0d0\ud5d8 \uc804\ub7b5\uc740 \uac01 \ud589\ub3d9\uc758 \uc120\ud0dd \ud655\ub960\uc744 \uadf8 \ud589\ub3d9\uc758 \uac00\uce58(value)\uc5d0 \ube44\ub840\ud558\uc5ec \uacc4\uc0b0\ud569\ub2c8\ub2e4. \uc774\ub294 \ub354 \ub192\uc740 \uac00\uce58\ub97c \uac00\uc9c4 \ud589\ub3d9\uc774 \ub354 \uc790\uc8fc \uc120\ud0dd\ub418\ub3c4\ub85d \ud558\uc5ec \ud0d0\ud5d8\uacfc \ud65c\uc6a9\uc758 \uade0\ud615\uc744 \ub9de\ucda5\ub2c8\ub2e4. \uc18c\ud504\ud2b8\ub9e5\uc2a4 \ud568\uc218\ub294 \uc544\ub798\uc640 \uac19\uc774 \uc815\uc758\ub429\ub2c8\ub2e4:<\/p>\n
def softmax(action_values, temperature=1.0):\n exp_values = np.exp(action_values \/ temperature)\n return exp_values \/ np.sum(exp_values)\n<\/code><\/pre>\n\uc5ec\uae30\uc11c \uc628\ub3c4 \ub9e4\uac1c\ubcc0\uc218\ub294 \ud0d0\ud5d8\uc758 \uac15\ub3c4\ub97c \uc870\uc808\ud569\ub2c8\ub2e4. \ub192\uc740 \uc628\ub3c4\uc5d0\uc11c\ub294 \uac01 \ud589\ub3d9\uc774 \ube44\uc2b7\ud55c \ud655\ub960\ub85c \uc120\ud0dd\ub418\ub294 \ubc18\uba74, \ub0ae\uc740 \uc628\ub3c4\uc5d0\uc11c\ub294 \uac00\uc7a5 \ub192\uc740 \uac00\uce58\uc758 \ud589\ub3d9\uc774 \uc120\ud0dd\ub420 \uac00\ub2a5\uc131\uc774 \ub192\uc544\uc9d1\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc18c\ud504\ud2b8\ub9e5\uc2a4 \ud0d0\ud5d8\uc744 \uad6c\ud604\ud558\ub294 \uc608\uc785\ub2c8\ub2e4:<\/p>\n
class Softmax:\n def __init__(self, n_actions, temperature=1.0):\n self.n_actions = n_actions\n self.temperature = temperature\n self.action_values = np.zeros(n_actions)\n\n def select_action(self):\n probabilities = softmax(self.action_values, self.temperature)\n return np.random.choice(self.n_actions, p=probabilities)\n\n def update(self, action, reward):\n self.action_values[action] += reward\n<\/code><\/pre>\n2.3 UCB (Upper Confidence Bound)<\/h3>\n
UCB\ub294 \uac01 \ud589\ub3d9\uc758 \uc2e0\ub8b0 \uad6c\uac04\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud0d0\ud5d8\uc744 \uc218\ud589\ud558\ub294 \uc804\ub7b5\uc785\ub2c8\ub2e4. \uc120\ud0dd\uc758 \ud69f\uc218\uc5d0 \ub530\ub77c \ubc1c\uc0dd\ud560 \uc218 \uc788\ub294 \ubd88\ud655\uc2e4\uc131\uc744 \uace0\ub824\ud558\uc5ec \ud589\ub3d9\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4. \uacf5\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
def ucb(action_values, action_counts, total_counts):\n return action_values + np.sqrt(2 * np.log(total_counts) \/ action_counts)\n<\/code><\/pre>\nUCB\uc758 \uacbd\uc6b0, \uc120\ud0dd\ub41c \ud589\ub3d9\uc758 \uc120\ud0dd \ud655\ub960\uc740 \uac01 \ud589\ub3d9\uc758 \uac00\uce58\uc640 \uadf8 \ud589\ub3d9\uc774 \uc120\ud0dd\ub41c \ud69f\uc218\uc5d0 \uc758\ud574 \uc870\uc815\ub429\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ud589\ub3d9\uc758 \ud655\uc2e4\uc131\uc744 \ub192\uc774\ub294 \ub3d9\uc2dc\uc5d0 \ud0d0\ud5d8\uc744 \uc7a5\ub824\ud569\ub2c8\ub2e4.<\/p>\n
class UCB:\n def __init__(self, n_actions):\n self.n_actions = n_actions\n self.action_values = np.zeros(n_actions)\n self.action_counts = np.zeros(n_actions)\n self.total_counts = 0\n\n def select_action(self):\n self.total_counts += 1\n ucb_values = ucb(self.action_values, self.action_counts, self.total_counts)\n return np.argmax(ucb_values)\n\n def update(self, action, reward):\n self.action_counts[action] += 1\n self.action_values[action] += (reward - self.action_values[action]) \/ self.action_counts[action]\n<\/code><\/pre>\n2.4 \ub2e4\ub978 \ud0d0\ud5d8 \uc804\ub7b5\ub4e4<\/h3>\n\n- Boltzmann Exploration: Boltzmann \ubd84\ud3ec\ub97c \uae30\ubc18\uc73c\ub85c \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\ub294 \ubc29\ubc95\uc73c\ub85c, \uc18c\ud504\ud2b8\ub9e5\uc2a4\uc640 \uc720\uc0ac.\n- Random Search: \ubb34\uc791\uc704\ub85c \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\ub294 \uae30\ubcf8\uc801\uc778 \ubc29\ubc95.\n- DECAY \u03b5-greedy: \uc2dc\uac04\uc5d0 \ub530\ub77c \u03b5 \uac12\uc744 \uac10\uc18c\uc2dc\ud0a4\ub294 \ubc29\ubc95\uc73c\ub85c, \ucd08\uae30\uc5d0\ub294 \ud0d0\ud5d8\uc744 \ub9ce\uc774 \ud558\ub2e4\uac00 \uc810\ucc28 \ud65c\uc6a9\uc744 \ub298\ub9ac\ub294 \uc804\ub7b5.\n\n\uac01 \ubc29\ubc95\uc740 \uc11c\ub85c \ub2e4\ub978 \uc0c1\ud669\uc5d0\uc11c \uc7a5\ub2e8\uc810\uc744 \uac00\uc9c0\uba70, \uc8fc\uc5b4\uc9c4 \ubb38\uc81c\uc758 \ud2b9\uc131\uc5d0 \ub530\ub77c \uc801\uc808\ud55c \ubc29\ubc95\uc744 \uc120\ud0dd\ud574\uc57c \ud569\ub2c8\ub2e4.\n<\/code><\/pre>\n3. \ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5 \ube44\uad50<\/h2>\n
\ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc120\ud0dd\ud558\uae30 \uc704\ud55c \uae30\uc900\uc740 \ubb38\uc81c\uc758 \ud2b9\uc131\uacfc \uc5d0\uc774\uc804\ud2b8\uc758 \ubaa9\ud45c\uc5d0 \ub530\ub77c \ub2e4\uc591\ud558\uac8c \ub098\ub258\uc5b4\uc9c8 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc5ec\ub7ec \uc804\ub7b5\uc758 \ube44\uad50\uc785\ub2c8\ub2e4:<\/p>\n
\n\n\uc804\ub7b5\uba85<\/th>\n \uc7a5\uc810<\/th>\n \ub2e8\uc810<\/th>\n<\/tr>\n \n\u03b5-\uadf8\ub9ac\ub514<\/td>\n \uac04\ub2e8\ud558\uace0 \uc774\ud574\ud558\uae30 \uc26c\uc6c0<\/td>\n \u03b5 \uac12 \uc124\uc815\uc774 \uc911\uc694\ud558\uba70, \u03b5 \uace0\uc815 \uc2dc \uc131\ub2a5 \uc800\ud558 \uac00\ub2a5<\/td>\n<\/tr>\n \n\uc18c\ud504\ud2b8\ub9e5\uc2a4<\/td>\n \uac01 \ud589\ub3d9\uc758 \uc120\ud0dd \ud655\ub960\uc744 \uae30\uc900\uc73c\ub85c \ud574\uc11c \ub354 \uc720\uc5f0\ud55c \ud0d0\ud5d8 \uac00\ub2a5<\/td>\n \uc628\ub3c4 \ub9e4\uac1c\ubcc0\uc218 \uc124\uc815\uc774 \uc694\uad6c\ub428<\/td>\n<\/tr>\n \nUCB<\/td>\n \uc2e0\ub8b0 \uad6c\uac04\uc744 \ubc18\uc601\ud558\uc5ec \ubcf4\ub2e4 \ud6a8\uc728\uc801\uc73c\ub85c \ud0d0\ud5d8 \uac00\ub2a5<\/td>\n \uacc4\uc0b0\uc774 \ubcf5\uc7a1\ud558\uba70, \uc804\ub9dd\uc774 \ubd88\ud655\uc2e4\ud55c \uc0c1\ud669\uc5d0\uc11c \ube44\ud6a8\uc728\uc801\uc77c \uc218 \uc788\uc74c<\/td>\n<\/tr>\n<\/table>\n4. \ud0d0\ud5d8 \uc804\ub7b5\uc758 \uc801\uc6a9 \ubc29\ubc95<\/h2>\n
\ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc801\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uacc4\uac00 \ud544\uc694\ud569\ub2c8\ub2e4:<\/p>\n
\n- \ud658\uacbd \uc815\uc758:<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \uc0c1\ud638\uc791\uc6a9\ud558\uac8c \ub420 \ud658\uacbd\uc744 \uc815\uc758\ud569\ub2c8\ub2e4.<\/li>\n
- \ud589\ub3d9 \uacf5\uac04 \ucd08\uae30\ud654:<\/strong> \uac01 \ud589\ub3d9\uc758 \uc120\ud0dd \ud655\ub960\uc774\ub098 \uac00\uce58 \ucd08\uae30\ud654\ub97c \ud1b5\ud574 \uc2dc\uc791\ud569\ub2c8\ub2e4.<\/li>\n
- \uc804\ub7b5 \uc120\ud0dd:<\/strong> \ubb38\uc81c\uc5d0 \uc801\ud569\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4.<\/li>\n
- \ud559\uc2b5 \ubc0f \uc5c5\ub370\uc774\ud2b8:<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba70 \uc5bb\uc740 \ubcf4\uc0c1\uc744 \uae30\ubc18\uc73c\ub85c \ud589\ub3d9\uc758 \uac00\uce58\ub97c \uc5c5\ub370\uc774\ud2b8 \ud569\ub2c8\ub2e4.<\/li>\n
- \uc131\ub2a5 \ud3c9\uac00:<\/strong> \ud559\uc2b5\uc774 \uc9c4\ud589\ub428\uc5d0 \ub530\ub77c \uc5d0\uc774\uc804\ud2b8\uc758 \uc131\uacfc\ub97c \uc815\uae30\uc801\uc73c\ub85c \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
5. \uacb0\ub860<\/h2>\n
\ud0d0\ud5d8\uacfc \ud65c\uc6a9\uc758 \uade0\ud615\uc744 \ub9de\ucd94\ub294 \uac83\uc740 \uac15\ud654\ud559\uc2b5\uc758 \uc131\uacf5\uc5d0 \ud544\uc218\uc801\uc785\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c \ub17c\uc758\ub41c \ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc740 \uac01\uae30 \ub2e4\ub978 \uc7a5\ub2e8\uc810\uc744 \uac00\uc9c0\uace0 \uc788\uc73c\uba70, \uc0c1\ud669\uc5d0 \ub9de\ucdb0 \uc801\uc808\ud788 \uc120\ud0dd\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \uac15\ud654\ud559\uc2b5 \ubd84\uc57c\ub294 \ub9e4\uc6b0 \ub113\uace0 \ubc1c\uc804 \uac00\ub2a5\uc131\uc774 \ud06c\ubbc0\ub85c, \uc774\ub97c \uc704\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc9c0\uc18d\uc801\uc73c\ub85c \uc5f0\uad6c\ud558\uace0 \uc801\uc6a9\ud558\ub294 \uac83\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\ub9c8\uc9c0\ub9c9\uc73c\ub85c, \uc774\ub860\uc744 \ubc14\ud0d5\uc73c\ub85c \ud55c \uc2e4\uc138\uacc4 \ubb38\uc81c \ud574\uacb0\uc744 \uc704\ud574\uc11c\ub294 \ub2e4\uc591\ud55c \ud0d0\ud5d8 \uc804\ub7b5\uc744 \uc9c0\uc18d\uc801\uc73c\ub85c \uc2dc\ud5d8\ud558\uace0 \ud3c9\uac00\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \uadf8\ub7ec\ubbc0\ub85c \uc5f0\uad6c\uc790\ub294 \uc774\ub7ec\ud55c \ud0d0\ud5d8 \uc804\ub7b5\ub4e4\uc744 \ub2e4\uc591\ud55c \ud658\uacbd\uc5d0\uc11c \uac80\uc99d\ud558\uace0 \ucd5c\uc801\ud654 \uacfc\uc815\uc744 \ubc18\ubcf5\ud558\uc5ec \ubcf4\ud3b8\uc801\uc778 \ud574\uacb0\ucc45\uc744 \ubc1c\uacac\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"