\uac15\ud654\ud559\uc2b5\uc740 \ubcf5\uc7a1\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uc0ac\uc6a9\ub418\uba70, \ub2e4\uc218\uc758 \ud658\uacbd\uacfc \ub2e4\uc591\ud55c \uc791\uc5c5\uc744 \ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc5d0\uc11c \ubcf5\uc7a1\uc131\uc774 \uc99d\uac00\ud569\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \ubaa8\ub4c8\ud654 \ubc0f \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720\ub97c \ud1b5\ud574 \uc774\ub7ec\ud55c \ubcf5\uc7a1\uc131\uc744 \uad00\ub9ac\ud558\uace0, \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd \uc811\uadfc \ubc29\uc2dd\uc744 \ud65c\uc6a9\ud558\uc5ec \ud6a8\uc728\uc801\uc778 \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ubaa8\ub4c8\ud654\ub780 \ud504\ub85c\uadf8\ub7a8\uc744 \uc5ec\ub7ec \uac1c\uc758 \ub3c5\ub9bd\uc801\uc778 \ucef4\ud3ec\ub10c\ud2b8\ub85c \ub098\ub204\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uac01 \ubaa8\ub4c8\uc740 \ub3c5\ub9bd\uc801\uc73c\ub85c \uc791\ub3d9\ud558\uba70, \ud2b9\uc815 \uae30\ub2a5\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. \ubaa8\ub4c8\ud654\uc758 \uc7a5\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\ub2e4\uc74c\uc740 \ud30c\uc774\uc36c\uc5d0\uc11c \ubaa8\ub4c8\ud654\ub97c \uc774\uc6a9\ud574 \uac15\ud654\ud559\uc2b5 \uc5d0\uc774\uc804\ud2b8\ub97c \uad6c\ud604\ud558\ub294 \uac04\ub2e8\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n
import numpy as np\n\nclass Environment:\n def __init__(self):\n self.state = 0\n\n def reset(self):\n self.state = 0\n return self.state\n\n def step(self, action):\n self.state += action\n reward = 1 if self.state >= 10 else 0\n return self.state, reward\n\nclass Agent:\n def __init__(self):\n self.q_table = np.zeros((10, 2))\n\n def choose_action(self, state):\n return np.argmax(self.q_table[state])\n\n def learn(self, state, action, reward, next_state):\n self.q_table[state, action] += 0.1 * (reward + 0.9 * np.max(self.q_table[next_state]) - self.q_table[state, action])\n\nenv = Environment()\nagent = Agent()\n\nfor episode in range(100):\n state = env.reset()\n done = False\n while not done:\n action = agent.choose_action(state)\n next_state, reward = env.step(action)\n agent.learn(state, action, reward, next_state)\n state = next_state\n if reward == 1:\n done = True<\/code><\/pre>\n3. \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720(Parameter Sharing)<\/h2>\n
\ud30c\ub77c\ubbf8\ud130 \uacf5\uc720\ub294 \uc5ec\ub7ec \uc791\uc5c5 \uac04\uc5d0 \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \uacf5\uc720\ud558\uc5ec \ud559\uc2b5 \ud6a8\uc728\uc131\uc744 \ub192\uc774\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774 \uae30\uc220\uc740 \ud2b9\ud788 \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc5d0\uc11c \uc720\uc6a9\ud569\ub2c8\ub2e4. \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720\uc758 \uc7a5\uc810\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ud6c8\ub828\uc5d0 \ud544\uc694\ud55c \ub370\uc774\ud130 \uc591\uc744 \uc904\uc77c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378\uc758 \uc77c\ubc18\ud654 \uc131\ub2a5\uc774 \ud5a5\uc0c1\ub429\ub2c8\ub2e4.<\/li>\n
- \ud6c8\ub828 \uc18d\ub3c4\uac00 \ube68\ub77c\uc9c0\uace0, \ud6a8\uc728\uc131\uc774 \ub192\uc544\uc9d1\ub2c8\ub2e4.<\/li>\n<\/ul>\n
3.1 \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720 \uc0ac\uc6a9 \uc608\uc81c<\/h3>\n
\uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc5d0\uc11c \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720\ub97c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c \uc608\uc81c\ub294 \ub450 \uac1c\uc758 \uacfc\uc81c\ub97c \uc218\ud589\ud558\ub294 \uc5d0\uc774\uc804\ud2b8\ub97c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n
class MultiTaskAgent:\n def __init__(self):\n self.q_table_1 = np.zeros((10, 2))\n self.q_table_2 = np.zeros((10, 2))\n\n def choose_action(self, state, task):\n if task == 1:\n return np.argmax(self.q_table_1[state])\n else:\n return np.argmax(self.q_table_2[state])\n\n def learn(self, state, action, reward, next_state, task):\n if task == 1:\n self.q_table_1[state, action] += 0.1 * (reward + 0.9 * np.max(self.q_table_1[next_state]) - self.q_table_1[state, action])\n else:\n self.q_table_2[state, action] += 0.1 * (reward + 0.9 * np.max(self.q_table_2[next_state]) - self.q_table_2[state, action])\n\n# \ub2e4\uc911 \uc5d0\uc774\uc804\ud2b8 \ud6c8\ub828\nagent = MultiTaskAgent()\nfor episode in range(100):\n for task in [1, 2]:\n state = env.reset()\n done = False\n while not done:\n action = agent.choose_action(state, task)\n next_state, reward = env.step(action)\n agent.learn(state, action, reward, next_state, task)\n state = next_state\n if reward == 1:\n done = True<\/code><\/pre>\n4. \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd<\/h2>\n
\uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd(Multi-task Learning)\uc740 \uc5ec\ub7ec \uc791\uc5c5\uc744 \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud558\ub294 \uae30\uacc4\ud559\uc2b5 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc5d0\uc11c\ub294 \ub2e4\uc591\ud55c \uc791\uc5c5\uc774 \uacf5\uc720\ud558\ub294 \uc815\ubcf4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \uc77c\ubc18\ud654 \ub2a5\ub825\uc744 \ud5a5\uc0c1\uc2dc\ud0b5\ub2c8\ub2e4.<\/p>\n
\uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc5d0\uc11c\ub294 \uac01 \uc791\uc5c5\uc758 \uace0\uc720\ud55c \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uba74\uc11c\ub3c4, \ubaa8\ub378\uc758 \uc77c\ubd80\ub97c \uacf5\uc720\ud558\uc5ec \ub370\uc774\ud130\uc758 \ud6a8\uc728\uc801\uc778 \uc0ac\uc6a9\uacfc \ube60\ub978 \ud559\uc2b5\uc744 \ub3c4\ubaa8\ud569\ub2c8\ub2e4.<\/p>\n
4.1 \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc758 \uc774\uc810<\/h3>\n\n- \ubaa8\ub378\uc758 \uc77c\ubc18\ud654 \ub2a5\ub825 \ud5a5\uc0c1<\/li>\n
- \ub370\uc774\ud130 \ud6a8\uc728\uc131 \uc99d\uac00<\/li>\n
- \ud559\uc2b5 \uc18d\ub3c4 \uc99d\uac00<\/li>\n<\/ul>\n
4.2 \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc744 \uc704\ud55c \uc544\ud0a4\ud14d\ucc98<\/h3>\n
\uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd \uc544\ud0a4\ud14d\ucc98\uc758 \uc77c\ubc18\uc801\uc778 \uc608\ub85c\ub294 \uacf5\uc720 \uc2e0\uacbd\ub9dd \uad6c\uc870\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uad6c\uc870\uc5d0\uc11c\ub294 \uae30\ubcf8\uc801\uc778 \uae30\ub2a5\uc744 \uc218\ud589\ud558\ub294 \uacf5\uc720 \ub808\uc774\uc5b4\uc640 \uac01 \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ud2b9\ud654\ub41c \ub808\uc774\uc5b4\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\n\nclass MultiTaskNetwork(nn.Module):\n def __init__(self):\n super(MultiTaskNetwork, self).__init__()\n self.shared_layer = nn.Linear(10, 5)\n self.task1_layer = nn.Linear(5, 2)\n self.task2_layer = nn.Linear(5, 2)\n\n def forward(self, x):\n shared_output = self.shared_layer(x)\n task1_output = self.task1_layer(shared_output)\n task2_output = self.task2_layer(shared_output)\n return task1_output, task2_output\n\nmodel = MultiTaskNetwork()\noptimizer = optim.Adam(model.parameters())\ncriterion = nn.MSELoss()\n\n# \ub370\uc774\ud130 \uc900\ube44 (\uac00\uc0c1\uc758 \ub370\uc774\ud130)\ndata1 = torch.randn(100, 10)\ndata2 = torch.randn(100, 10)\nlabels1 = torch.randint(0, 2, (100, 2)).float()\nlabels2 = torch.randint(0, 2, (100, 2)).float()\n\nfor epoch in range(100):\n model.train()\n optimizer.zero_grad()\n \n # task 1\n outputs1, outputs2 = model(data1)\n loss1 = criterion(outputs1, labels1)\n \n # task 2\n outputs1, outputs2 = model(data2)\n loss2 = criterion(outputs2, labels2)\n \n # \uc804\uccb4 \uc190\uc2e4\n loss = loss1 + loss2\n loss.backward()\n optimizer.step()<\/code><\/pre>\n5. \uc2e4\uc6a9\uc801\uc778 \ud301<\/h2>\n\n- \ubaa8\ub4c8\ud654 \uc2dc \ucd94\uc0c1\ud654\ub97c \uc801\uc808\ud788 \ud65c\uc6a9\ud558\uc138\uc694.<\/li>\n
- \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720 \uc2dc \uc5b4\ub5a4 \ud30c\ub77c\ubbf8\ud130\ub97c \uacf5\uc720\ud560\uc9c0 \uc2e0\uc911\ud788 \uacb0\uc815\ud558\uc138\uc694.<\/li>\n
- \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd\uc5d0\uc11c\ub294 \ubaa8\ub4e0 \uc791\uc5c5 \uac04\uc758 \uc5f0\uad00\uc131\uc744 \uace0\ub824\ud558\uc138\uc694.<\/li>\n
- \uc2e4\ud5d8\uc744 \ud1b5\ud574 \ub2e4\uc591\ud55c \uad6c\uc870\ub97c \ube44\uad50\ud558\uace0 \uc131\ub2a5\uc744 \ubd84\uc11d\ud558\uc138\uc694.<\/li>\n<\/ul>\n
6. \uacb0\ub860<\/h2>\n
\ubaa8\ub4c8\ud654\uc640 \ud30c\ub77c\ubbf8\ud130 \uacf5\uc720\ub294 \ubcf5\uc7a1\ud55c \uac15\ud654\ud559\uc2b5 \ubb38\uc81c\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \uad00\ub9ac\ud558\ub294 \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\ubc95\uc744 \ud65c\uc6a9\ud558\uc5ec \uba40\ud2f0\ud0dc\uc2a4\ud06c \ub7ec\ub2dd \ud658\uacbd\uc5d0\uc11c\ub3c4 \ud6a8\uc728\uc801\uc778 \ubaa8\ub378\uc744 \uc124\uacc4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ube14\ub85c\uadf8 \ud3ec\uc2a4\ud2b8\uac00 \uac15\ud654\ud559\uc2b5 \uac1c\ubc1c\uc790\ub4e4\uc5d0\uac8c \uc720\uc775\ud55c \uc815\ubcf4\uac00 \ub418\uc5c8\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub2e4\uc591\ud55c \uac15\uc88c\ub97c \ud1b5\ud574 \ub354\uc6b1 \uc2ec\ud654\ub41c \ub0b4\uc6a9\uc744 \ub2e4\ub8f0 \uc608\uc815\uc785\ub2c8\ub2e4.<\/p>\n
7. \ucc38\uace0 \ubb38\ud5cc<\/h2>\n