\ucd5c\uadfc \uc778\uacf5\uc9c0\ub2a5(AI) \ubc0f \uba38\uc2e0\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c\ub294 \ub2e8\uc77c \uc5d0\uc774\uc804\ud2b8\ub97c \ub118\uc5b4 \uc5ec\ub7ec \uc5d0\uc774\uc804\ud2b8\uac00 \uc11c\ub85c \ud611\ub825\ud558\uac70\ub098 \uacbd\uc7c1\ud558\ub294 \uc0c1\ud669\uc5d0\uc11c \ud559\uc2b5\ud558\ub294 \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5(Multi-Agent Reinforcement Learning, MARL)\uc5d0 \ub300\ud55c \uad00\uc2ec\uc774 \uc99d\uac00\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc740 \uc5ec\ub7ec \uc5d0\uc774\uc804\ud2b8\uac00 \uc0c1\ud638\uc791\uc6a9\ud558\ub294 \ud658\uacbd\uc5d0\uc11c \ucd5c\uc801\uc758 \ud589\ub3d9 \uc804\ub7b5\uc744 \ucc3e\ub294 \ubb38\uc81c\ub97c \ub2e4\ub8f9\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc758 \uc774\ub860\uacfc \uc2e4\uc81c \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc124\uba85\ud558\uace0, \ud611\ub825\uacfc \uacbd\uc7c1\uc758 \uc5d0\uc774\uc804\ud2b8 \uc804\ub7b5\uc5d0 \ub300\ud55c \uc608\uc81c \uc18c\uc2a4\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
\uac15\ud654\ud559\uc2b5(Reinforcement Learning, RL)\uc740 \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba70 \ubcf4\uc0c1\uc744 \uadf9\ub300\ud654\ud558\ub294 \ud589\ub3d9\uc744 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc5d0\uc11c\ub294 \uc5ec\uae30\uc11c \uc815\uc758\ub41c \uc5d0\uc774\uc804\ud2b8\ub4e4\uc774 \ub354\ud574\uc838 \uc5ec\ub7ec \uc5d0\uc774\uc804\ud2b8\uac00 \uc874\uc7ac\ud558\ub294 \ud658\uacbd\uc744 \uad6c\uc131\ud569\ub2c8\ub2e4. \uac01 \uc5d0\uc774\uc804\ud2b8\ub294 \ub2e4\ub978 \uc5d0\uc774\uc804\ud2b8\uc640 \uacbd\uc7c1\ud558\uac70\ub098 \ud611\ub825\ud558\uc5ec \ucd5c\uc801\uc758 \uc804\ub7b5\uc744 \ubc30\uc6cc\ub098\uac11\ub2c8\ub2e4.<\/p>\n
\uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \ud658\uacbd\uc740 \ud06c\uac8c \ub450 \uac00\uc9c0\ub85c \ub098\ub20c \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc5d0\ub294 \uc5ec\ub7ec \uac00\uc9c0 \uc911\uc694\ud55c \uad6c\uc131 \uc694\uc18c\uac00 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc740 \ub9ce\uc740 \ub3c4\uc804 \uacfc\uc81c\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\ud611\ub825\uc801 \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc5d0\uc11c\ub294 \uc5d0\uc774\uc804\ud2b8\ub4e4\uc774 \uacf5\ub3d9\uc758 \ubaa9\ud45c\ub97c \ub2ec\uc131\ud558\uae30 \uc704\ud574 \ud611\ub825\ud558\ub294 \uc804\ub7b5\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \ub2e4\uc591\ud55c \uae30\ubc95\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4:<\/p>\n
\nimport numpy as np\nimport random\n\nclass Environment:\n def __init__(self, num_agents):\n self.num_agents = num_agents\n self.state = np.zeros(num_agents) # \uc5d0\uc774\uc804\ud2b8 \uc0c1\ud0dc \ucd08\uae30\ud654\n self.goal = np.array([1] * num_agents) # \uacf5\ub3d9 \ubaa9\ud45c\n\n def step(self, actions):\n rewards = []\n for i in range(self.num_agents):\n self.state[i] += actions[i] # \uac01 \uc5d0\uc774\uc804\ud2b8\uc758 \ud589\ub3d9\uc744 \uc0c1\ud0dc\uc5d0 \ubc18\uc601\n if self.state[i] >= self.goal[i]:\n rewards.append(1) # \ubaa9\ud45c \ub2ec\uc131 \ubcf4\uc0c1\n else:\n rewards.append(-0.1) # \ubaa9\ud45c \ubbf8\ub2ec\uc131 \ud398\ub110\ud2f0\n return rewards\n\nclass Agent:\n def __init__(self, id):\n self.id = id\n self.policy = np.random.rand(2) # \uc815\ucc45 \ucd08\uae30\ud654 (\uc774\uc0b0 \ud589\ub3d9)\n \n def select_action(self, state):\n return np.argmax(np.dot(state, self.policy)) # \ud589\ub3d9 \uc120\ud0dd\n\nnum_agents = 3\nenv = Environment(num_agents)\nagents = [Agent(i) for i in range(num_agents)]\n\nfor episode in range(100): # \ud6c8\ub828 \uc5d0\ud53c\uc18c\ub4dc\n state = env.state\n actions = [agent.select_action(state) for agent in agents] # \uac01 \uc5d0\uc774\uc804\ud2b8\uc758 \ud589\ub3d9 \uc120\ud0dd\n rewards = env.step(actions) # \ud658\uacbd\uc5d0\uc11c \ub2e8\uacc4 \uc9c4\ud589\n<\/code><\/code><\/pre>\n5. \uacbd\uc7c1\uc801 MARL\uc744 \uc704\ud55c \uae30\ubc95<\/h2>\n
\uacbd\uc7c1\uc801 \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc5d0\uc11c\ub294 \uac01 \uc5d0\uc774\uc804\ud2b8\uac00 \uc790\uc2e0\ub9cc\uc758 \uc774\uc775\uc744 \uadf9\ub300\ud654\ud558\ub824\uace0 \ud569\ub2c8\ub2e4. \uc774\uc5d0 \ub530\ub77c \ub2e4\uc74c\uacfc \uac19\uc740 \uc804\ub7b5\ub4e4\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4:<\/p>\n
\n- Stackelberg \uac8c\uc784:<\/strong> \ud558\ub098\uc758 \uc120\ub3c4 \uc5d0\uc774\uc804\ud2b8\uc640 \uc5ec\ub7ec \ucd94\uc885 \uc5d0\uc774\uc804\ud2b8\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\ub294 \uac8c\uc784 \uc774\ub860 \ubaa8\ub378\uc785\ub2c8\ub2e4. \uc120\ub3c4 \uc5d0\uc774\uc804\ud2b8\uac00 \ud589\ub3d9\uc744 \uba3c\uc800 \uacb0\uc815\ud55c \ud6c4, \ucd94\uc885 \uc5d0\uc774\uc804\ud2b8\ub4e4\uc774 \uc774\uc5d0 \ub530\ub77c \ubc18\uc751\ud569\ub2c8\ub2e4.<\/li>\n
- \ub098\uc26c \uade0\ud615\uacc4(Nash Equilibrium):<\/strong> \uac01 \uc5d0\uc774\uc804\ud2b8\uac00 \uc790\uc2e0\uc758 \ucd5c\uc801 \uc804\ub7b5\uc744 \uc120\ud0dd\ud560 \ub54c \ub2e4\ub978 \uc5d0\uc774\uc804\ud2b8\uc758 \ud589\ub3d9\uc5d0 \ubcc0\ud654\uac00 \uc5c6\ub3c4\ub85d \ud558\ub294 \uc0c1\ud0dc\uc785\ub2c8\ub2e4. \uc774 \uc0c1\ud0dc\uc5d0\uc11c \uac01 \uc5d0\uc774\uc804\ud2b8\ub294 \ucd5c\uc801\uc758 \uacb0\uc815\uc744 \ub0b4\ub9bd\ub2c8\ub2e4.<\/li>\n<\/ul>\n
5.1 \uacbd\uc7c1\uc801 MARL \uc608\uc81c<\/h3>\n\nclass CompetitiveEnvironment:\n def __init__(self):\n self.state = 0 # \ucd08\uae30 \uc0c1\ud0dc\n \n def step(self, actions):\n reward = []\n for action in actions:\n reward.append(action - self.state) # \ud589\ub3d9\uacfc \ud604\uc7ac \uc0c1\ud0dc\uc5d0 \ub530\ub978 \ubcf4\uc0c1\n self.state += action # \ud589\ub3d9 \ubc18\uc601\n return reward\n\nclass CompetitiveAgent:\n def __init__(self, id):\n self.id = id\n self.q_table = np.zeros(10) # Q-\ud14c\uc774\ube14 \ucd08\uae30\ud654\n self.exploration_rate = 0.1\n\n def select_action(self):\n if random.random() < self.exploration_rate:\n return random.choice(range(10)) # \ud0d0\ud5d8\n return np.argmax(self.q_table) # \ucd5c\uc801 \ud589\ub3d9 \uc120\ud0dd\n\nenv = CompetitiveEnvironment()\nagents = [CompetitiveAgent(i) for i in range(2)]\n\nfor episode in range(100): # \ud6c8\ub828 \uc5d0\ud53c\uc18c\ub4dc\n actions = [agent.select_action() for agent in agents] # \uac01 \uc5d0\uc774\uc804\ud2b8\uc758 \ud589\ub3d9 \uc120\ud0dd\n rewards = env.step(actions) # \ud658\uacbd\uc5d0\uc11c \ub2e8\uacc4 \uc9c4\ud589\n for i, reward in enumerate(rewards):\n # Q-\ud14c\uc774\ube14 \uc5c5\ub370\uc774\ud2b8 (\uac04\ub2e8\ud55c \ud615\ud0dc)\n agents[i].q_table[actions[i]] += 0.1 * (reward - agents[i].q_table[actions[i]])\n<\/code><\/code><\/pre>\n6. MARL\uc758 \ud5a5\ud6c4 \ubc29\ud5a5<\/h2>\n
\uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc758 \uc5f0\uad6c\ub294 \uc5ec\uc804\ud788 \ud65c\ubc1c\ud788 \uc9c4\ud589 \uc911\uc774\uba70, \ud611\ub825\uc801 \ubc0f \uacbd\uc7c1\uc801 \uc0c1\ud669 \ubaa8\ub450\uc5d0\uc11c \uc5d0\uc774\uc804\ud2b8\ub4e4\uc774 \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ub2e4\uc591\ud55c \uc811\uadfc\ubc95\uc774 \ubaa8\uc0c9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uc5f0\ud569 \ud559\uc2b5(Federated Learning)\uacfc \uac19\uc740 \uc0c8\ub85c\uc6b4 \ud559\uc2b5 \uae30\ubc95\ub4e4\uc774 \uc801\uc6a9\ub420 \uac00\ub2a5\uc131\uc774 \uc788\uc73c\uba70, \uc774\ub85c \uc778\ud574 \uc5d0\uc774\uc804\ud2b8\uac00 \uc11c\ub85c\uc758 \uacbd\ud5d8\uc744 \uacf5\uc720\ud558\uba74\uc11c\ub3c4 \uac1c\uc778\uc758 \ube44\ubc00\uc744 \ubcf4\ud638\ud560 \uc218 \uc788\ub294 \ubc1c\uc804\uc774 \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n
7. \uacb0\ub860<\/h2>\n
\uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc740 \uc2e4\uc81c \uc138\uacc4\uc758 \ubcf5\uc7a1\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \uc788\uc5b4 \uc911\uc694\ud55c \ub3c4\uad6c\ub85c \uc790\ub9ac \uc7a1\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud611\ub825\uc801\uc774\ub4e0 \uacbd\uc7c1\uc801\uc774\ub4e0 \uc5d0\uc774\uc804\ud2b8\ub4e4\uc774 \uc0c1\ud638\uc791\uc6a9\ud558\ub294 \ud658\uacbd\uc5d0\uc11c \ud6a8\uacfc\uc801\uc778 \uc815\ucc45\uc744 \ud559\uc2b5\ud558\ub294 \uac83\uc740 AI\uc758 \ubbf8\ub798 \ubc1c\uc804\uc5d0 \uc788\uc5b4 \ub9e4\uc6b0 \uc911\uc694\ud569\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c \uc18c\uac1c\ud55c \uc774\ub860\uacfc \uac04\ub2e8\ud55c \uc608\uc81c\ub294 \uba40\ud2f0\uc5d0\uc774\uc804\ud2b8 \uac15\ud654\ud559\uc2b5\uc5d0 \ub300\ud55c \uae30\ucd08\uc801 \uc774\ud574\ub97c \ub3d5\uae30 \uc704\ud55c \uac83\uc785\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ub354 \ub9ce\uc740 \uc5f0\uad6c\uc640 \uc0ac\ub840\uac00 \ud544\uc694\ud558\uba70, \ub3c5\uc790 \uc5ec\ub7ec\ubd84\uc774 \uc774 \ubd84\uc57c\uc5d0\uc11c \uc0c8\ub85c\uc6b4 \uac00\ub2a5\uc131\uc744 \ud0d0\uc0c9\ud560 \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n