Gym Documentation<\/a><\/p>\nHumanoid\u793a\u4f8b\u5982\u4e0b\uff1a<\/strong><\/p>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926171232160.gif\")
\n<\/p>\n
cart_pole\u793a\u4f8b\u5982\u4e0b\uff1a<\/strong><\/p>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926171256629.gif\")
\n<\/p>\n
\u672c\u9879\u76ee\u4e5f\u662f\u5728\u6b64\u6e38\u620f\u73af\u5883\u4e0b\u8fdb\u884c\u7684\u4ee3\u7801\u793a\u4f8b\u3002<\/p>\n
![](\"https:\/\/img.icons8.com\/?size=100&id=48250&format=png&color=000000\")
\u793a\u4f8b\u4ee3\u7801\uff1a\u57fa\u4e8e\u4ef7\u503c\u7684\u5f3a\u5316\u5b66\u4e60<\/h3>\n
\nimport torch \nimport torch.nn as nn \nimport torch.optim as optim \nimport torch.nn.functional as F \nimport numpy as np \nimport random \nimport math \nimport gym\n\n# \u5b9a\u4e49\u91cd\u653e\u7f13\u51b2\u533a\u7c7b\uff0c\u7528\u4e8e\u5b58\u50a8\u7ecf\u9a8c\nclass ReplayBuffer:\n # \u521d\u59cb\u5316\u65b9\u6cd5\uff0c\u5b9a\u4e49\u7f13\u51b2\u533a\u5bb9\u91cf\u548c\u4f4d\u7f6e\n def __init__(self, capacity):\n self.buffer = [] \n self.capacity = capacity \n self.position = 0 \n\n # \u5411\u7f13\u51b2\u533a\u6dfb\u52a0\u7ecf\u9a8c\n def push(self, transition):\n if len(self.buffer) < self.capacity: \n self.buffer.append(None) \n self.buffer[self.position] = transition \n self.position = (self.position + 1) % self.capacity \n\n # \u4ece\u7f13\u51b2\u533a\u968f\u673a\u91c7\u6837\u4e00\u4e2a\u6279\u6b21\u7684\u7ecf\u9a8c\n def sample(self, batch_size):\n batch = random.sample(self.buffer, batch_size) \n state, action, reward, next_state, done = map(np.stack, zip(*batch))\n return state, action, reward, next_state, done\n\n # \u8fd4\u56de\u7f13\u51b2\u533a\u7684\u5f53\u524d\u5927\u5c0f\n def __len__(self):\n return len(self.buffer)\n\n# \u5b9a\u4e49\u7b56\u7565\u7f51\u7edc\u7c7b\nclass PolicyNetwork(nn.Module):\n\n def __init__(self, n_states, n_actions):\n super(PolicyNetwork, self).__init__()\n self.fc1 = nn.Linear(n_states, 128) \n self.fc2 = nn.Linear(128, n_actions) \n\n def forward(self, x):\n x = F.relu(self.fc1(x)) \n x = self.fc2(x) \n return x \n\n# \u5b9a\u4e49\u6df1\u5ea6Q\u7f51\u7edc\u7c7b\nclass DQN:\n # \u521d\u59cb\u5316\u65b9\u6cd5\uff0c\u5b9a\u4e49\u76f8\u5173\u53c2\u6570\u548c\u7f51\u7edc\n def __init__(self, model, memory, cfg):\n self.n_actions = cfg['n_actions'] # \u52a8\u4f5c\u6570\u91cf\n self.device = torch.device(cfg['device']) # \u8bbe\u5907\uff08CPU\u6216GPU\uff09\n self.gamma = cfg['gamma'] # \u6298\u6263\u56e0\u5b50\n self.sample_count = 0 # \u91c7\u6837\u8ba1\u6570\n self.epsilon = cfg['epsilon_start'] # \u521d\u59cbepsilon\u503c\n self.epsilon_start = cfg['epsilon_start'] # epsilon\u521d\u59cb\u503c\n self.epsilon_end = cfg['epsilon_end'] # epsilon\u6700\u7ec8\u503c\n self.epsilon_decay = cfg['epsilon_decay'] # epsilon\u8870\u51cf\u7387\n self.batch_size = cfg['batch_size'] # \u6279\u91cf\u5927\u5c0f\n self.policy_net = model.to(self.device) # \u7b56\u7565\u7f51\u7edc\n self.target_net = model.to(self.device) # \u76ee\u6807\u7f51\u7edc\n # \u521d\u59cb\u5316\u76ee\u6807\u7f51\u7edc\u6743\u91cd\u4e0e\u7b56\u7565\u7f51\u7edc\u76f8\u540c\n for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()): \n target_param.data.copy_(param.data)\n self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg['lr']) # \u4f18\u5316\u5668\n self.memory = memory # \u91cd\u653e\u7f13\u51b2\u533a\n\n # \u91c7\u6837\u52a8\u4f5c\u65b9\u6cd5\uff0c\u57fa\u4e8eepsilon-greedy\u7b56\u7565\n def sample_action(self, state):\n self.sample_count += 1 # \u589e\u52a0\u91c7\u6837\u8ba1\u6570\n # \u66f4\u65b0epsilon\u503c\uff0c\u91c7\u7528\u6307\u6570\u8870\u51cf\n self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \\\n math.exp(-1. * self.sample_count \/ self.epsilon_decay) \n if random.random() > self.epsilon: # \u9009\u62e9\u8d2a\u5a6a\u7b56\u7565\n with torch.no_grad():\n state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(dim=0)\n q_values = self.policy_net(state) # \u8ba1\u7b97Q\u503c\n action = q_values.max(1)[1].item() # \u9009\u62e9Q\u503c\u6700\u5927\u7684\u52a8\u4f5c\n else: # \u968f\u673a\u9009\u62e9\u52a8\u4f5c\n action = random.randrange(self.n_actions)\n return action # \u8fd4\u56de\u52a8\u4f5c\n\n # \u9884\u6d4b\u52a8\u4f5c\u65b9\u6cd5\uff0c\u7528\u4e8e\u6d4b\u8bd5\u9636\u6bb5\n @torch.no_grad()\n def predict_action(self, state):\n state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(dim=0)\n q_values = self.policy_net(state) # \u8ba1\u7b97Q\u503c\n action = q_values.max(1)[1].item() # \u9009\u62e9Q\u503c\u6700\u5927\u7684\u52a8\u4f5c\n return action # \u8fd4\u56de\u52a8\u4f5c\n\n # \u66f4\u65b0\u7f51\u7edc\u65b9\u6cd5\n def update(self):\n if len(self.memory) < self.batch_size: # \u5982\u679c\u7f13\u51b2\u533a\u4e2d\u7684\u7ecf\u9a8c\u4e0d\u8db3\u4e00\u4e2a\u6279\u6b21\n return\n # \u4ece\u7f13\u51b2\u533a\u91c7\u6837\u4e00\u4e2a\u6279\u6b21\u7684\u7ecf\u9a8c\n state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)\n # \u8f6c\u6362\u4e3a\u5f20\u91cf\n state_batch = torch.tensor(np.array(state_batch), device=self.device, dtype=torch.float)\n action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)\n reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float)\n next_state_batch = torch.tensor(np.array(next_state_batch), device=self.device, dtype=torch.float)\n done_batch = torch.tensor(np.float32(done_batch), device=self.device)\n # \u8ba1\u7b97\u5f53\u524dQ\u503c\n q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch)\n # \u8ba1\u7b97\u4e0b\u4e00\u4e2a\u72b6\u6001\u7684Q\u503c\n next_q_values = self.target_net(next_state_batch).max(1)[0].detach()\n # \u8ba1\u7b97\u671f\u671bQ\u503c\n expected_q_values = reward_batch + self.gamma * next_q_values * (1 - done_batch)\n # \u8ba1\u7b97\u635f\u5931\n loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))\n self.optimizer.zero_grad() # \u6e05\u7a7a\u68af\u5ea6\n loss.backward() # \u53cd\u5411\u4f20\u64ad\n # \u68af\u5ea6\u88c1\u526a\n for param in self.policy_net.parameters():\n param.grad.data.clamp_(-1, 1)\n self.optimizer.step() # \u66f4\u65b0\u7f51\u7edc\u53c2\u6570\n\n# \u73af\u5883\u548c\u667a\u80fd\u4f53\u914d\u7f6e\u51fd\u6570\ndef env_agent_config(cfg):\n env = gym.make(cfg['env_name']) # \u521b\u5efa\u73af\u5883\n n_states = env.observation_space.shape[0] # \u72b6\u6001\u7a7a\u95f4\u7ef4\u5ea6\n n_actions = env.action_space.n # \u52a8\u4f5c\u7a7a\u95f4\u7ef4\u5ea6\n memory = ReplayBuffer(cfg['memory_capacity']) # \u521b\u5efa\u91cd\u653e\u7f13\u51b2\u533a\n agent = DQN(PolicyNetwork(n_states, n_actions), memory, cfg) # \u521b\u5efaDQN\u667a\u80fd\u4f53\n return env, agent # \u8fd4\u56de\u73af\u5883\u548c\u667a\u80fd\u4f53\n\n# \u8bad\u7ec3\u51fd\u6570\ndef train(cfg, env, agent):\n print("\u5f00\u59cb\u8bad\u7ec3\uff01")\n rewards = [] # \u5b58\u50a8\u6bcf\u56de\u5408\u7684\u5956\u52b1\n steps = [] # \u5b58\u50a8\u6bcf\u56de\u5408\u7684\u6b65\u6570\n for i_ep in range(cfg['train_eps']):\n ep_reward = 0 # \u521d\u59cb\u5316\u56de\u5408\u5956\u52b1\n ep_step = 0 # \u521d\u59cb\u5316\u56de\u5408\u6b65\u6570\n state, _ = env.reset() # \u91cd\u7f6e\u73af\u5883\uff0c\u83b7\u53d6\u521d\u59cb\u72b6\u6001\n for _ in range(cfg['ep_max_steps']):\n ep_step += 1\n action = agent.sample_action(state) # \u91c7\u6837\u52a8\u4f5c\n next_state, reward, done, truncated, info = env.step(action) # \u6267\u884c\u52a8\u4f5c\n done = done or truncated\n agent.memory.push((state, action, reward, next_state, done)) # \u5c06\u7ecf\u9a8c\u5b58\u50a8\u5230\u7f13\u51b2\u533a\n state = next_state # \u66f4\u65b0\u72b6\u6001\n agent.update() # \u66f4\u65b0\u8bc4\u4f30\u7f51\u7edc\n ep_reward += reward # \u7d2f\u52a0\u5956\u52b1\n if done: # \u5982\u679c\u56de\u5408\u7ed3\u675f\n break\n if (i_ep + 1) % cfg['target_update'] == 0: # \u6bcf\u9694\u4e00\u5b9a\u56de\u5408\u66f4\u65b0\u76ee\u6807\u7f51\u7edc\n agent.target_net.load_state_dict(agent.policy_net.state_dict())\n steps.append(ep_step) # \u8bb0\u5f55\u6b65\u6570\n rewards.append(ep_reward) # \u8bb0\u5f55\u5956\u52b1\n if (i_ep + 1) % 10 == 0:\n print(f"\u56de\u5408\uff1a{i_ep + 1}\/{cfg['train_eps']}\uff0c\u5956\u52b1\uff1a{ep_reward:.2f}\uff0cEpislon\uff1a{agent.epsilon:.3f}")\n print("\u5b8c\u6210\u8bad\u7ec3\uff01")\n env.close() # \u5173\u95ed\u73af\u5883\n return {'rewards': rewards} # \u8fd4\u56de\u5956\u52b1\u8bb0\u5f55\n\n# \u6d4b\u8bd5\u51fd\u6570\ndef test(cfg, env, agent):\n print("\u5f00\u59cb\u6d4b\u8bd5\uff01")\n rewards = [] # \u5b58\u50a8\u6bcf\u56de\u5408\u7684\u5956\u52b1\n steps = [] # \u5b58\u50a8\u6bcf\u56de\u5408\u7684\u6b65\u6570\n for i_ep in range(cfg['test_eps']):\n ep_reward = 0 # \u521d\u59cb\u5316\u56de\u5408\u5956\u52b1\n ep_step = 0 # \u521d\u59cb\u5316\u56de\u5408\u6b65\u6570\n state, _ = env.reset() # \u91cd\u7f6e\u73af\u5883\uff0c\u83b7\u53d6\u521d\u59cb\u72b6\u6001\n done = False # \u521d\u59cb\u5316\u7ed3\u675f\u6807\u5fd7\n while not done and ep_step < cfg['ep_max_steps']:\n ep_step += 1\n action = agent.predict_action(state) # \u9884\u6d4b\u52a8\u4f5c\n next_state, reward, done, truncated, info = env.step(action) # \u6267\u884c\u52a8\u4f5c\n done = done or truncated\n state = next_state # \u66f4\u65b0\u72b6\u6001\n ep_reward += reward # \u7d2f\u52a0\u5956\u52b1\n if done:\n print(f"\u56de\u5408 {i_ep + 1} \u5728\u6b65\u6570 {ep_step} \u5b8c\u6210")\n steps.append(ep_step) # \u8bb0\u5f55\u6b65\u6570\n rewards.append(ep_reward) # \u8bb0\u5f55\u5956\u52b1\n print(f"\u56de\u5408\uff1a{i_ep + 1}\/{cfg['test_eps']}\uff0c\u5956\u52b1\uff1a{ep_reward:.2f}")\n print("\u5b8c\u6210\u6d4b\u8bd5")\n env.close() # \u5173\u95ed\u73af\u5883\n return {'rewards': rewards} # \u8fd4\u56de\u5956\u52b1\u8bb0\u5f55\n\n# \u83b7\u53d6\u914d\u7f6e\u53c2\u6570\u51fd\u6570\ndef get_args():\n return {\n 'env_name': 'CartPole-v1', # \u73af\u5883\u540d\u79f0\n 'n_actions': 2, # \u52a8\u4f5c\u6570\u91cf\n 'device': 'cpu', # \u8bbe\u5907\uff08CPU\uff09\n 'gamma': 0.99, # \u6298\u6263\u56e0\u5b50\n 'epsilon_start': 1.0, # \u521d\u59cbepsilon\u503c\n 'epsilon_end': 0.01, # \u6700\u7ec8epsilon\u503c\n 'epsilon_decay': 500, # epsilon\u8870\u51cf\u7387\n 'batch_size': 64, # \u6279\u91cf\u5927\u5c0f\n 'lr': 0.001, # \u5b66\u4e60\u7387\n 'memory_capacity': 10000, # \u7f13\u51b2\u533a\u5bb9\u91cf\n 'train_eps': 100, # \u8bad\u7ec3\u56de\u5408\u6570\n 'test_eps': 100, # \u6d4b\u8bd5\u56de\u5408\u6570\n 'ep_max_steps': 200, # \u6bcf\u56de\u5408\u6700\u5927\u6b65\u6570\n 'target_update': 10, # \u76ee\u6807\u7f51\u7edc\u66f4\u65b0\u9891\u7387\n 'seed': 42 # \u968f\u673a\u79cd\u5b50\n }\n\n# \u7ed8\u5236\u5956\u52b1\u66f2\u7ebf\u51fd\u6570\ndef plot_rewards(rewards, cfg, tag="train"):\n import matplotlib.pyplot as plt\n plt.plot(rewards) # \u7ed8\u5236\u5956\u52b1\u66f2\u7ebf\n plt.title(f'{tag} Rewards') # \u8bbe\u7f6e\u6807\u9898\n plt.show() # \u663e\u793a\u56fe\u50cf\n\n# \u83b7\u53d6\u53c2\u6570\ncfg = get_args()\n# \u8bad\u7ec3\nenv, agent = env_agent_config(cfg)\nres_dic = train(cfg, env, agent)\n\n# \u7ed8\u5236\u8bad\u7ec3\u5956\u52b1\u66f2\u7ebf\nplot_rewards(res_dic['rewards'], cfg, tag="train")\n\n# \u6d4b\u8bd5\nres_dic = test(cfg, env, agent)\n\n# \u7ed8\u5236\u6d4b\u8bd5\u5956\u52b1\u66f2\u7ebf\nplot_rewards(res_dic['rewards'], cfg, tag="test")\n<\/code><\/pre>\n\u5f00\u59cb\u8bad\u7ec3\uff01\n\u56de\u5408\uff1a10\/100\uff0c\u5956\u52b1\uff1a25.00\uff0cEpislon\uff1a0.698\n\u56de\u5408\uff1a20\/100\uff0c\u5956\u52b1\uff1a16.00\uff0cEpislon\uff1a0.517\n\u56de\u5408\uff1a30\/100\uff0c\u5956\u52b1\uff1a12.00\uff0cEpislon\uff1a0.366\n\u56de\u5408\uff1a40\/100\uff0c\u5956\u52b1\uff1a78.00\uff0cEpislon\uff1a0.130\n\u56de\u5408\uff1a50\/100\uff0c\u5956\u52b1\uff1a64.00\uff0cEpislon\uff1a0.045\n\u56de\u5408\uff1a60\/100\uff0c\u5956\u52b1\uff1a52.00\uff0cEpislon\uff1a0.022\n\u56de\u5408\uff1a70\/100\uff0c\u5956\u52b1\uff1a47.00\uff0cEpislon\uff1a0.014\n\u56de\u5408\uff1a80\/100\uff0c\u5956\u52b1\uff1a44.00\uff0cEpislon\uff1a0.011\n\u56de\u5408\uff1a90\/100\uff0c\u5956\u52b1\uff1a115.00\uff0cEpislon\uff1a0.010\n\u56de\u5408\uff1a100\/100\uff0c\u5956\u52b1\uff1a200.00\uff0cEpislon\uff1a0.010\n\u5b8c\u6210\u8bad\u7ec3\uff01<\/code><\/pre>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926171629841.png\")
\n<\/p>\n
\u5f00\u59cb\u6d4b\u8bd5\uff01\n\u56de\u5408\uff1a1\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a2\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a3\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a4\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a5\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a6\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a7\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a8\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a9\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a10\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408 11 \u5728\u6b65\u6570 115 \u5b8c\u6210\n\u56de\u5408\uff1a11\/100\uff0c\u5956\u52b1\uff1a115.00\n\u56de\u5408\uff1a12\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a13\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a14\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a15\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a16\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a17\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a18\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a19\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a20\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a21\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a22\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a23\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a24\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a25\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a26\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a27\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a28\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a29\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a30\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a31\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a32\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a33\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a34\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a35\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a36\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a37\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a38\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a39\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a40\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a41\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a42\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a43\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a44\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a45\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a46\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a47\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a48\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a49\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a50\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a51\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a52\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a53\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a54\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a55\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a56\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a57\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a58\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a59\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a60\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a61\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408 62 \u5728\u6b65\u6570 118 \u5b8c\u6210\n\u56de\u5408\uff1a62\/100\uff0c\u5956\u52b1\uff1a118.00\n\u56de\u5408\uff1a63\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a64\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408 65 \u5728\u6b65\u6570 113 \u5b8c\u6210\n\u56de\u5408\uff1a65\/100\uff0c\u5956\u52b1\uff1a113.00\n\u56de\u5408\uff1a66\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a67\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a68\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a69\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a70\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a71\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a72\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a73\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a74\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a75\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a76\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a77\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a78\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a79\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a80\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a81\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a82\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a83\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a84\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a85\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408 86 \u5728\u6b65\u6570 109 \u5b8c\u6210\n\u56de\u5408\uff1a86\/100\uff0c\u5956\u52b1\uff1a109.00\n\u56de\u5408 87 \u5728\u6b65\u6570 123 \u5b8c\u6210\n\u56de\u5408\uff1a87\/100\uff0c\u5956\u52b1\uff1a123.00\n\u56de\u5408\uff1a88\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a89\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a90\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a91\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a92\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a93\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a94\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a95\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a96\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a97\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a98\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a99\/100\uff0c\u5956\u52b1\uff1a200.00\n\u56de\u5408\uff1a100\/100\uff0c\u5956\u52b1\uff1a200.00\n\u5b8c\u6210\u6d4b\u8bd5<\/code><\/pre>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926171655404.png\")
\n<\/p>\n
![](\"https:\/\/img.icons8.com\/?size=100&id=xaquNfre75yC&format=png&color=000000\")
\u57fa\u4e8e\u7b56\u7565\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\uff08Policy-based model\uff09<\/h2>\n
\n\n- \u60f3\u8c61\u4e00\u4e0b\uff0c\u4f60\u6b63\u5728\u6559\u4e00\u4e2a\u673a\u5668\u4eba\u5b66\u4e60\u5982\u4f55\u8d70\u8def\u3002<\/li>\n
- \u5728\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u4e2d\uff0c\u4f60\u76f4\u63a5\u544a\u8bc9\u8fd9\u4e2a\u673a\u5668\u4eba\u5728\u6bcf\u4e00\u6b65\u8be5\u5982\u4f55\u884c\u52a8\u3002\u8fd9\u79cd\u6307\u5bfc\u662f\u901a\u8fc7\u4e00\u4e2a\u6982\u7387\u6a21\u578b\u6765\u5b9e\u73b0\u7684\uff0c\u5373\u7b56\u7565\u51fd\u6570\u3002<\/li>\n
- \u7b56\u7565\u51fd\u6570\u8003\u8651\u5f53\u524d\u7684\u73af\u5883\uff08\u5373\u673a\u5668\u4eba\u7684\u5f53\u524d\u72b6\u6001\uff09\uff0c\u7136\u540e\u8f93\u51fa\u6bcf\u4e00\u4e2a\u53ef\u80fd\u52a8\u4f5c\u7684\u6982\u7387\u3002\u673a\u5668\u4eba\u6839\u636e\u8fd9\u4e9b\u6982\u7387\u6765\u9009\u62e9\u5176\u52a8\u4f5c\u3002<\/li>\n
- \u56e0\u6b64\uff0c\u5728\u6bcf\u4e2a\u72b6\u6001\u4e0b\uff0c\u673a\u5668\u4eba\u6709\u4e00\u4e2a\u6e05\u6670\u7684\u6982\u7387\u6307\u5bfc\u5b83\u51b3\u5b9a\u4e0b\u4e00\u6b65\u5982\u4f55\u884c\u52a8\u3002<\/li>\n
- \u8fd9\u79cd\u65b9\u6cd5\u7684\u5173\u952e\u4f18\u52bf\u662f\u53ef\u4ee5\u5904\u7406\u66f4\u52a0\u590d\u6742\u548c\u591a\u6837\u7684\u52a8\u4f5c\u9009\u62e9\uff0c\u4f8b\u5982\u5728\u8fde\u7eed\u52a8\u4f5c\u7a7a\u95f4\u4e2d\u9009\u62e9\u5408\u9002\u7684\u52a8\u4f5c\u529b\u5ea6\u548c\u65b9\u5411\u3002<\/li>\n<\/ul>\n
\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5728DQN\u8fd9\u7c7b\u57fa\u4e8e\u503c\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u4e2d\uff0c\u6211\u4eec\u4e0d\u662f\u76f4\u63a5\u544a\u8bc9\u673a\u5668\u4eba\u6bcf\u4e00\u6b65\u600e\u4e48\u505a\uff0c\u800c\u662f\u7ed9\u5b83\u4e00\u4e2a\u8bc4\u5206\u7cfb\u7edf<\/strong>\uff1a\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u53ef\u80fd\u7684\u52a8\u4f5c\uff0cDQN\u8bc4\u4f30\u5176\u5e26\u6765\u7684\u9884\u671f\u56de\u62a5\uff08\u5373Q\u503c\uff09\u3002\u673a\u5668\u4eba\u7684\u4efb\u52a1\u662f\u5728\u6bcf\u4e00\u6b65\u9009\u62e9\u90a3\u4e2a\u9884\u671f\u56de\u62a5\u6700\u9ad8\u7684\u52a8\u4f5c\u3002\u8fd9\u5c31\u597d\u6bd4\u4e0d\u662f\u544a\u8bc9\u673a\u5668\u4eba\u5177\u4f53\u600e\u4e48\u8d70\uff0c\u800c\u662f\u7ed9\u6bcf\u4e00\u6b65\u4e00\u4e2a\u5206\u6570\uff0c\u8ba9\u5b83\u81ea\u5df1\u627e\u51fa\u5f97\u5206\u6700\u9ad8\u7684\u6b65\u9aa4\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u79bb\u6563\u52a8\u4f5c\u7a7a\u95f4<\/strong>\u4e2d\u8868\u73b0\u5f97\u975e\u5e38\u597d\uff0c\u56e0\u4e3a\u53ef\u4ee5\u8f7b\u677e\u6bd4\u8f83\u548c\u9009\u62e9\u6709\u9650\u4e2a\u52a8\u4f5c\u4e2d\u7684\u6700\u4f73\u9009\u9879\u3002<\/p>\n\u73b0\u5728\uff0c\u5982\u679c\u52a8\u4f5c\u7a7a\u95f4\u53d8\u5f97\u975e\u5e38\u5927\u6216\u8005\u662f\u8fde\u7eed\u7684<\/strong>\uff08\u60f3\u8c61\u8ba9\u673a\u5668\u4eba\u5b66\u4e60\u600e\u6837\u4ee5\u4efb\u610f\u901f\u5ea6\u548c\u65b9\u5411\u8dd1\u6b65\uff09\uff0c\u57fa\u4e8e\u503c\u7684\u65b9\u6cd5\u5c31\u53d8\u5f97\u590d\u6742\u4e14\u6548\u7387\u4e0d\u9ad8\u4e86\u3002\u56e0\u4e3a\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u53ef\u80fd\u7684\u52a8\u4f5c\uff0c\u4f60\u90fd\u9700\u8981\u8ba1\u7b97\u4e00\u4e2a\u9884\u671f\u56de\u62a5\uff0c\u5e76\u4ece\u4e2d\u9009\u62e9\u6700\u4f73\u7684\u3002\u4f46\u5728\u8fde\u7eed\u52a8\u4f5c\u7a7a\u95f4\u4e2d\uff0c\u53ef\u80fd\u7684\u52a8\u4f5c\u662f\u65e0\u9650\u7684\uff0c\u8fd9\u4f7f\u5f97\u627e\u51fa\u6700\u4f73\u52a8\u4f5c\u53d8\u5f97\u975e\u5e38\u56f0\u96be\u3002<\/p>\n\u76f8\u53cd\uff0c\u57fa\u4e8e\u7b56\u7565\u7684\u65b9\u6cd5\u5219\u53ef\u4ee5\u66f4\u81ea\u7136\u5730\u5904\u7406\u8fd9\u79cd\u60c5\u51b5\u3002\u56e0\u4e3a\u4f60\u4e0d\u9700\u8981\u4e3a\u6bcf\u4e2a\u53ef\u80fd\u7684\u52a8\u4f5c\u5206\u522b\u8ba1\u7b97\u4e00\u4e2a\u503c\uff0c\u800c\u662f\u76f4\u63a5\u6839\u636e\u5f53\u524d\u72b6\u6001\u751f\u6210\u52a8\u4f5c\u7684\u6982\u7387\u5206\u5e03\u3002\u8fd9\u4f7f\u5f97\u673a\u5668\u4eba\u80fd\u591f\u5728\u8fde\u7eed\u7684\u52a8\u4f5c\u7a7a\u95f4\u4e2d\u5e73\u6ed1\u5730\u9009\u62e9\u52a8\u4f5c\uff0c\u800c\u4e0d\u662f\u5728\u6709\u9650\u7684\u9009\u9879\u4e2d\u505a\u51fa\u9009\u62e9\u3002<\/p>\n
\u7b56\u7565\u68af\u5ea6\u4f18\u5316<\/h3>\n\n- \n
\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u76f4\u63a5\u5bf9\u7b56\u7565\u672c\u8eab\u8fdb\u884c\u4f18\u5316\uff0c\u800c\u4e0d\u662f\u50cf\u5728DQN\u4e2d\u90a3\u6837\u95f4\u63a5\u901a\u8fc7Q\u503c\u6765\u63a8\u5bfc\u7b56\u7565\u3002<\/p>\n<\/li>\n
- \n
\u7b56\u7565\u672c\u8eab\u901a\u5e38\u7531\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u8868\u793a\uff0c\u8f93\u5165\u662f\u72b6\u6001\uff0c\u8f93\u51fa\u662f\u5728\u8fd9\u4e2a\u72b6\u6001\u4e0b\u91c7\u53d6\u6bcf\u4e2a\u53ef\u80fd\u52a8\u4f5c\u7684\u6982\u7387\u3002<\/p>\n<\/li>\n
- \n
\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u662f\u901a\u8fc7\u8c03\u6574\u7b56\u7565\u7f51\u7edc\u7684\u53c2\u6570\u6765\u76f4\u63a5\u6700\u5927\u5316\u7d2f\u79ef\u56de\u62a5\u7684\u671f\u671b\u503c\u3002<\/p>\n<\/li>\n
- \n
\u5176\u4e2d\uff0cREINFORCE \u7b97\u6cd5<\/strong>\u662f\u4e00\u79cd\u7ecf\u5178\u7684\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u3002<\/p>\n<\/li>\n- \n
\u5728REINFORCE\u4e2d\uff0c\u6211\u4eec\u8ba9\u7b56\u7565\u7f51\u7edc\u6307\u5bfc\u52a8\u4f5c\u7684\u9009\u62e9\uff0c\u7136\u540e\u6839\u636e\u5b9e\u9645\u83b7\u5f97\u7684\u56de\u62a5\u6765\u8c03\u6574\u7f51\u7edc\u53c2\u6570\u3002<\/p>\n<\/li>\n
- \n
\u7b80\u77ed\u8bf4\uff0c\u5982\u679c\u4e00\u4e2a\u52a8\u4f5c\u5bfc\u81f4\u4e86\u6b63\u7684\u56de\u62a5\uff0c\u90a3\u4e48\u6211\u4eec\u5c31\u589e\u52a0\u672a\u6765\u9009\u62e9\u8fd9\u4e2a\u52a8\u4f5c\u7684\u6982\u7387\uff1b\u53cd\u4e4b\uff0c\u5982\u679c\u5bfc\u81f4\u4e86\u8d1f\u9762\u7684\u56de\u62a5\uff0c\u90a3\u4e48\u9009\u62e9\u8fd9\u4e2a\u52a8\u4f5c\u7684\u6982\u7387\u5c31\u4f1a\u51cf\u5c11\u3002<\/p>\n<\/li>\n
- \n
\u5177\u4f53\u7684\u7b97\u6cd5\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n
\n- \n
\u521d\u59cb\u5316\u7b56\u7565\u6a21\u578b<\/strong>\uff1a\u7b56\u7565\u901a\u5e38\u7531\u4e00\u4e2a\u53c2\u6570\u5316\u7684\u6a21\u578b\u8868\u793a\uff0c\u6bd4\u5982\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u3002\u8fd9\u4e2a\u7f51\u7edc\u6839\u636e\u8f93\u5165\u7684\u72b6\u6001\u6765\u8f93\u51fa\u6bcf\u4e2a\u53ef\u884c\u52a8\u4f5c\u7684\u6982\u7387\u3002<\/p>\n<\/li>\n- \n
\u751f\u6210\u4e00\u4e2a\u6216\u591a\u4e2a\u7247\u6bb5<\/strong>\uff1a\u5728\u6bcf\u4e2a\u8fed\u4ee3\u4e2d\uff0c\u4f7f\u7528\u5f53\u524d\u7684\u7b56\u7565\u5728\u73af\u5883\u4e2d\u6267\u884c\u591a\u4e2a\u5b8c\u6574\u7684\u8f68\u8ff9\uff08Trajectory\uff09\uff0c\u76f4\u5230\u8fbe\u5230\u7ec8\u6b62\u72b6\u6001\u3002\u6bcf\u4e2a\u8f68\u8ff9\u662f\u4e00\u4e2a\u72b6\u6001\u3001\u52a8\u4f5c\u548c\u56de\u62a5\u7684\u5e8f\u5217\u3002<\/p>\n<\/li>\n- \n
\u8ba1\u7b97\u56de\u62a5<\/strong>\uff1a\u5bf9\u4e8e\u8f68\u8ff9\u4e2d\u7684\u6bcf\u4e2a\u6b65\u9aa4\uff0c\u8ba1\u7b97\u4ece\u5f53\u524d\u72b6\u6001\u5230\u8f68\u8ff9\u7ed3\u675f\u6240\u83b7\u5f97\u7684\u7d2f\u79ef\u56de\u62a5\u3002\u8fd9\u4e2a\u7d2f\u79ef\u56de\u62a5\u901a\u5e38\u662f\u5bf9\u672a\u6765\u56de\u62a5\u7684\u4e00\u4e2a\u6298\u6263\u7d2f\u8ba1\u548c\uff0c\u5176\u4e2d\u6298\u6263\u56e0\u5b50 \u03b3 \u51b3\u5b9a\u4e86\u672a\u6765\u56de\u62a5\u7684\u6743\u91cd\u3002<\/p>\n<\/li>\n- \n
\u7b56\u7565\u68af\u5ea6\u66f4\u65b0<\/strong>\uff1a\u5bf9\u4e8e\u8f68\u8ff9\u4e2d\u7684\u6bcf\u4e2a\u6b65\u9aa4\uff0c\u8c03\u6574\u7b56\u7565\u6a21\u578b\u7684\u53c2\u6570\uff0c\u4ee5\u589e\u52a0\u5728\u7ed9\u5b9a\u72b6\u6001\u4e0b\u9009\u62e9\u5b9e\u9645\u6267\u884c\u7684\u52a8\u4f5c\u7684\u6982\u7387\u3002\u8fd9\u4e2a\u8c03\u6574\u662f\u6839\u636e\u7b56\u7565\u68af\u5ea6\u8fdb\u884c\u7684\uff0c\u901a\u5e38\u4f7f\u7528\u68af\u5ea6\u4e0a\u5347\u6cd5\u6765\u5b9e\u73b0\u3002\u7b56\u7565\u68af\u5ea6\u7531\u6bcf\u4e2a\u6b65\u9aa4\u7684\u7d2f\u79ef\u56de\u62a5\u548c\u52a8\u4f5c\u6982\u7387\u7684\u68af\u5ea6\u7684\u4e58\u79ef\u7ed9\u51fa\u3002 \u91cd\u590d\u4e0a\u8ff0\u6b65\u9aa4\u76f4\u5230\u7b56\u7565\u6536\u655b\u6216\u8fbe\u5230\u67d0\u4e2a\u6027\u80fd\u6807\u51c6\u3002<\/p>\n<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n\u7b56\u7565\u68af\u5ea6\u7684\u57fa\u672c\u5f62\u5f0f\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<\/p>\n
$$
\n\\nabla_\\theta J(\\theta)=\\mathbb{E}\\left[\\sum_{t=0}^T G_t \\cdot \\nabla_\\theta \\log \\pi_\\theta\\left(a_t \\mid s_t\\right)\\right]
\n$$<\/p>\n
\u4ece\u76f4\u89c2\u4e0a\u6211\u4eec\u53ef\u4ee5\u8fd9\u4e9b\u89e3\u91ca\u4e0a\u8ff0\u516c\u5f0f\uff1a<\/p>\n
\n- \u671f\u671b\u56de\u62a5<\/strong>: $J(\\theta)$ \u662f\u5728\u7b56\u7565 $\\pi_\\theta$ \u4e0b\u7684\u671f\u671b\u56de\u62a5\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u8c03\u6574 $\\theta$ \u6765\u6700\u5927\u5316 $J(\\theta)$ \u3002<\/li>\n
- \u68af\u5ea6\u4e0a\u5347<\/strong>: \u7531\u4e8e\u6211\u4eec\u8981\u6700\u5927\u5316 $J(\\theta)$ \uff0c\u6211\u4eec\u9700\u8981\u671d\u7740\u68af\u5ea6 $\\nabla_\\theta J(\\theta)$ \u7684\u65b9\u5411\u8c03\u6574 $\\theta$ \u3002\u8fd9\u610f\u5473\u7740\u5982\u679c\u67d0\u4e2a\u65b9\u5411\u4e0a\u7684\u68af\u5ea6\u4e3a\u6b63\uff0c\u589e\u52a0 $\\theta$ \u5728\u8fd9\u4e2a\u65b9\u5411\u4e0a\u7684\u503c\u4f1a\u589e\u52a0\u671f\u671b\u56de\u62a5\u3002<\/li>\n
- \u7b56\u7565\u6982\u7387\u7684\u5bf9\u6570<\/strong>: $\\log \\pi_\\theta\\left(a_t \\mid s_t\\right)$ \u662f\u5728\u7ed9\u5b9a\u72b6\u6001 $s_t$ \u4e0b\u9009\u62e9\u52a8\u4f5c $a_t$ \u7684\u7b56\u7565\u6982\u7387\u7684\u81ea\u7136\u5bf9\u6570\u3002\u5bf9\u6570\u8fd0\u7b97\u4f7f\u5f97\u68af\u5ea6\u8ba1\u7b97\u66f4\u4e3a\u7a33\u5b9a\u548c\u9ad8\u6548\u3002<\/li>\n
- \u68af\u5ea6<\/strong> $\\nabla_\\theta \\log \\pi_\\theta\\left(a_t \\mid s_t\\right)$ : \u8fd9\u90e8\u5206\u8868\u660e\u7b56\u7565\u6982\u7387\u5bf9\u53c2\u6570 $\\theta$ \u7684\u654f\u611f\u7a0b\u5ea6\u3002\u76f4\u89c2\u4e0a\u8bf4\uff0c\u5b83\u544a\u8bc9\u6211\u4eec\uff0c\u5fae\u5c0f\u5730\u8c03\u6574 $\\theta$\u4f1a\u5982\u4f55\u5f71\u54cd\u5728\u72b6\u6001 $s_t$ \u4e0b\u9009\u62e9\u52a8\u4f5c $a_t$ \u7684\u6982\u7387\u3002<\/li>\n
- \u7d2f\u79ef\u56de\u62a5<\/strong> $G_t$ : \u8fd9\u662f\u4ece\u65f6\u95f4 $t$ \u5230\u8f68\u8ff9\u7ed3\u675f\u7684\u7d2f\u79ef\u6298\u6263\u56de\u62a5\u3002 $G_t$ \u7684\u5f15\u5165\u786e\u4fdd\u4e86\u52a8\u4f5c\u7684\u9009\u62e9\u4e0d\u4ec5\u8003\u8651\u7acb\u5373\u56de\u62a5\uff0c\u800c\u662f\u57fa\u4e8e\u5176\u957f\u671f\u5f71\u54cd\u3002\u901a\u8fc7 $G_t$ \uff0c\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u53ef\u4ee5\u4fc3\u8fdb\u90a3\u4e9b\u53ef\u80fd\u7acb\u5373\u56de\u62a5\u4e0d\u9ad8\uff0c\u4f46\u957f\u671f\u6548\u679c\u597d\u7684\u52a8\u4f5c\u3002<\/li>\n
- \u671f\u671b<\/strong>: \u5916\u5c42\u7684\u671f\u671b $\\mathbb{E}$ \u8868\u793a\u8fd9\u4e2a\u68af\u5ea6\u662f\u5728\u6240\u6709\u53ef\u80fd\u8f68\u8ff9\u4e0b\u7684\u5e73\u5747\u68af\u5ea6\u3002\u7531\u4e8e\u76f4\u63a5\u8ba1\u7b97\u6574\u4e2a\u671f\u671b\u901a\u5e38\u4e0d\u53ef\u884c\uff0c\u5b9e\u8df5\u4e2d\u5e38\u5e38\u901a\u8fc7\u4ece\u5f53\u524d\u7b56\u7565\u4e2d\u91c7\u6837\u4e00\u4e2a\u6216\u591a\u4e2a\u8f68\u8ff9\u6765\u4f30\u8ba1\u8fd9\u4e2a\u671f\u671b\u3002<\/li>\n<\/ul>\n
![](\"https:\/\/img.icons8.com\/?size=100&id=91CnU00i6HLv&format=png&color=000000\")
\u4e3a\u4ec0\u4e48\u8981\u5c06\u7b56\u7565\u6982\u7387\u5bf9\u6570\u7684\u68af\u5ea6\u4e0e\u7d2f\u79ef\u56de\u62a5\u76f8\u4e58?<\/h4>\n
\n\u5bf9\u6bd4\u68af\u5ea6\u4e0a\u5347\u548c\u65f6\u5e8f\u5dee\u5206<\/h3>\n
\u68af\u5ea6\u4e0a\u5347\u548c\u65f6\u5e8f\u5dee\u5206\u5b66\u4e60\uff08TD-Learning\uff09\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u4e24\u79cd\u4e0d\u540c\u7684\u4f18\u5316\u7b56\u7565\uff0c\u5b83\u4eec\u5404\u6709\u4f18\u52bf\u548c\u5c40\u9650\u6027\u3002\u4e0b\u9762\u5bf9\u8fd9\u4e24\u79cd\u65b9\u6cd5\u8fdb\u884c\u5bf9\u6bd4\uff1a<\/p>\n
\u5728\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u4e2d\u4f7f\u7528\u68af\u5ea6\u4e0a\u5347\u7684\u4f18\u70b9\u4e3b\u8981\u4f53\u73b0\u4e0b\u4ee5\u4e0b\u4e09\u65b9\u9762\uff1a<\/p>\n
\n- \n
\u76f4\u63a5\u4f18\u5316\u7b56\u7565\uff1a\u68af\u5ea6\u4e0a\u5347\u76f4\u63a5\u5bf9\u7b56\u7565\u8fdb\u884c\u4f18\u5316\uff0c\u4f7f\u5f97\u5b66\u4e60\u8fc7\u7a0b\u66f4\u76f4\u63a5\u548c\u9ad8\u6548\uff0c\u7279\u522b\u662f\u5728\u590d\u6742\u6216\u8fde\u7eed\u7684\u52a8\u4f5c\u7a7a\u95f4\u4e2d\u3002<\/p>\n<\/li>\n
- \n
\u66f4\u597d\u7684\u63a2\u7d22\uff1a\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u901a\u5e38\u80fd\u66f4\u597d\u5730\u63a2\u7d22\u52a8\u4f5c\u7a7a\u95f4\uff0c\u56e0\u4e3a\u5b83\u4eec\u901a\u8fc7\u6982\u7387\u6027\u7b56\u7565\u6765\u9009\u62e9\u52a8\u4f5c\uff0c\u8fd9\u610f\u5473\u7740\u603b\u6709\u4e00\u5b9a\u7684\u6982\u7387\u53bb\u5c1d\u8bd5\u4e0d\u540c\u7684\u52a8\u4f5c\uff0c\u800c\u4e0d\u662f\u4e00\u76f4\u9009\u62e9\u6700\u5927\u7684\u6982\u7387\u52a8\u4f5c\u3002<\/p>\n<\/li>\n
- \n
\u907f\u514d\u5c40\u90e8\u6700\u4f18\uff1a\u7531\u4e8e\u5176\u63a2\u7d22\u7279\u6027\u548c\u76f4\u63a5\u4f18\u5316\u7b56\u7565\u7684\u80fd\u529b\uff0c\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u53ef\u80fd\u66f4\u5bb9\u6613\u8df3\u51fa\u5c40\u90e8\u6700\u4f18\u89e3\u3002<\/p>\n<\/li>\n<\/ol>\n
\u5728\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u4e2d\u4f7f\u7528\u68af\u5ea6\u4e0a\u5347\u7684\u7f3a\u70b9\u5982\u4e0b\uff1a<\/p>\n
\n- \n
\u9ad8\u65b9\u5dee\uff1a\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u7279\u522b\u662f\u57fa\u4e8e\u8499\u7279\u5361\u6d1b\u7684\u7b56\u7565\u68af\u5ea6\uff0c\u53ef\u80fd\u4f1a\u53d7\u5230\u8f83\u9ad8\u65b9\u5dee\u7684\u5f71\u54cd\uff0c\u4ece\u800c\u5f71\u54cd\u5b66\u4e60\u7684\u7a33\u5b9a\u6027\u548c\u6536\u655b\u901f\u5ea6\u3002<\/p>\n<\/li>\n
- \n
\u6837\u672c\u6548\u7387\u4f4e\uff1a\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u9700\u8981\u66f4\u591a\u7684\u6837\u672c\u6765\u51c6\u786e\u4f30\u8ba1\u68af\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u590d\u6742\u73af\u5883\u4e2d\u3002<\/p>\n<\/li>\n
- \n
\u8c03\u53c2\u654f\u611f\uff1a\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u5728\u5b9e\u8df5\u4e2d\u53ef\u80fd\u66f4\u4f9d\u8d56\u4e8e\u8c03\u53c2\uff08\u5982\u5b66\u4e60\u7387\u3001\u7b56\u7565\u7f51\u7edc\u7684\u7ed3\u6784\u7b49\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n
\u5728DQN\u4e2d\u4f7f\u7528\u65f6\u5e8f\u5dee\u5206\u5b66\u4e60\u7684\u4f18\u70b9\u5982\u4e0b\uff1a<\/p>\n
\n- \n
\u6837\u672c\u6548\u7387\u9ad8\uff1aTD-Learning\u901a\u8fc7\u4f7f\u7528\u6bcf\u4e2a\u6b65\u9aa4\u7684\u66f4\u65b0\uff0c\u800c\u4e0d\u662f\u7b49\u5f85\u6574\u4e2a\u8f68\u8ff9\u5b8c\u6210\uff0c\u901a\u5e38\u66f4\u52a0\u6837\u672c\u6548\u7387\u9ad8\u3002<\/p>\n<\/li>\n
- \n
\u7a33\u5b9a\u6027\uff1a\u76f8\u6bd4\u4e8e\u8499\u7279\u5361\u6d1b\u65b9\u6cd5\u7684\u7b56\u7565\u68af\u5ea6\uff0cTD-Learning\u901a\u5e38\u80fd\u63d0\u4f9b\u66f4\u7a33\u5b9a\u7684\u5b66\u4e60\u66f4\u65b0\uff0c\u56e0\u4e3aQ\u503c\u7684\u66f4\u65b0\u4e0d\u5b8c\u5168\u4f9d\u8d56\u4e8e\u5b9e\u9645\u83b7\u5f97\u7684\u957f\u671f\u56de\u62a5\uff0c\u800c\u662f\u90e8\u5206\u5730\u4f9d\u8d56\u4e8e\u5bf9\u672a\u6765\u4ef7\u503c\u7684\u73b0\u6709\u4f30\u8ba1\uff0c\u52a0\u4e0a\u5f53\u524d\u72b6\u6001\u7684\u89c2\u6d4b\u4ef7\u503c\uff0c\u4ece\u800c\u4f7f\u5b66\u4e60\u66f4\u65b0\u66f4\u52a0\u5e73\u6ed1\u548c\u7a33\u5b9a\u3002<\/p>\n<\/li>\n
- \n
\u4f4e\u65b9\u5dee\uff1a\u65b9\u5dee\u8fd9\u91cc\u6307\u7684\u662f\u4ece\u4e00\u4e2a\u66f4\u65b0\u5230\u4e0b\u4e00\u4e2a\u66f4\u65b0\u4f30\u8ba1\u503c\u53d8\u5316\u7684\u4e0d\u786e\u5b9a\u6027\u3002\u5728\u8499\u7279\u5361\u6d1b\u65b9\u6cd5\u4e2d\uff0c\u503c\u51fd\u6570\u7684\u6bcf\u6b21\u66f4\u65b0\u662f\u5b8c\u5168\u57fa\u4e8e\u6574\u4e2a\u8f68\u8ff9\u7684\u5b9e\u9645\u56de\u62a5\uff0c\u8fd9\u5bfc\u81f4\u4e86\u4f30\u8ba1\u7684\u65b9\u5dee\u8f83\u9ad8\uff0c\u5c24\u5176\u662f\u5728\u56de\u62a5\u4fe1\u53f7\u53d8\u5316\u8f83\u5927\u6216\u8f68\u8ff9\u5f88\u957f\u65f6\u3002\u800c\u5728TD-Learning\u4e2d\uff0c\u7531\u4e8e\u6bcf\u6b21\u66f4\u65b0\u90e8\u5206\u57fa\u4e8e\u5f53\u524d\u7684\u503c\u51fd\u6570\u4f30\u8ba1\uff0c\u8fd9\u610f\u5473\u7740\u6bcf\u4e2a\u66f4\u65b0\u66f4\u52a0\u5e73\u7a33\uff0c\u5bf9\u4e8e\u5355\u4e00\u5f02\u5e38\u7684\u72b6\u6001\u8f6c\u6362\u6216\u56de\u62a5\u4e0d\u4f1a\u8fc7\u5ea6\u53cd\u5e94\uff0c\u4ece\u800c\u5bfc\u81f4\u6574\u4f53\u7684\u4f30\u8ba1\u65b9\u5dee\u8f83\u4f4e\u3002<\/p>\n<\/li>\n<\/ol>\n
\u5728DQN\u4e2d\u4f7f\u7528\u65f6\u5e8f\u5dee\u5206\u5b66\u4e60\u7684\u7f3a\u70b9\u5982\u4e0b\uff1a<\/p>\n
\n- \n
\u5728TD-Learning\u4e2d\uff0c\u201d\u53cc\u91cd\u5229\u7528\u201d (double sampling) \u95ee\u9898\u6307\u7684\u662f\u5728\u540c\u4e00\u8fc7\u7a0b\u4e2d\u5bf9\u540c\u4e00\u4e2a\u6837\u672c\u6570\u636e (\u6bd4\u5982\u72b6\u6001\u8f6c\u79fb) \u7684\u91cd\u590d\u4f7f\u7528\u3002\u8fd9\u79cd\u81ea\u5df1\u6293\u81ea\u5df1\u5c3e\u5df4\u7684\u884c\u4e3a\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u5bf9\u4e8e\u672a\u6765\u56de\u62a5\u7684\u8fc7\u5ea6\u4e50\u89c2\u4f30\u8ba1\uff0c\u8fdb\u800c\u5bfc\u81f4\u5b66\u4e60\u7684\u504f\u5dee\uff0c\u5f53\u7136\uff0c\u5728\u540e\u7eed\u7684DQN\u6539\u8fdb\u7b97\u6cd5\u4e2d\uff0c\u8fd9\u4e2a\u53cc\u91cd\u5229\u7528\u7684\u95ee\u9898\u53ef\u4ee5\u901a\u8fc7\u5212\u5206\u8bc4\u4ef7\u7f51\u7edc\u548c\u76ee\u6807\u7f51\u7edc\u6765\u7f13\u89e3\u3002<\/p>\n<\/li>\n
- \n
\u63a2\u7d22\u95ee\u9898\u6d89\u53ca\u5230\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\uff1a\u5728\u57fa\u4e8e\u503c\u7684\u65b9\u6cd5\u5982DQN\u4e2d\uff0c\u901a\u5e38\u4f7f\u7528\u8d2a\u5fc3\u7b56\u7565\uff08\u603b\u662f\u9009\u62e9\u5f53\u524d\u4f30\u8ba1\u503c\u6700\u9ad8\u7684\u52a8\u4f5c\uff09\u6216 \u03f5-\u8d2a\u5fc3\u7b56\u7565\uff08\u5927\u90e8\u5206\u65f6\u95f4\u9009\u62e9\u6700\u4f18\u52a8\u4f5c\uff0c\u4f46\u6709\u4e00\u4e2a\u5c0f\u6982\u7387 \u03f5 \u968f\u673a\u9009\u62e9\u5176\u4ed6\u52a8\u4f5c\uff09\u6765\u5e73\u8861\u8fd9\u4e24\u8005\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u80fd\u4e0d\u591f\u7075\u6d3b\uff0c\u7279\u522b\u662f\u5728\u52a8\u4f5c\u7a7a\u95f4\u5f88\u5927\u6216\u8fde\u7eed\u7684\u73af\u5883\u4e2d\uff0c\u56e0\u4e3a\u5b83\u4eec\u53ef\u80fd\u4f1a\u5ffd\u89c6\u90a3\u4e9b\u77ed\u671f\u5185\u770b\u8d77\u6765\u4e0d\u662f\u6700\u4f18\uff0c\u4f46\u957f\u671f\u6765\u770b\u53ef\u80fd\u66f4\u597d\u7684\u7b56\u7565\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u6982\u7387\u6027\u7b56\u7565\uff08\u6bd4\u5982\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u4e2d\u4f7f\u7528\u7684\uff09\u5141\u8bb8\u7b97\u6cd5\u5728\u6bcf\u4e2a\u72b6\u6001\u4e0b\u6839\u636e\u67d0\u79cd\u6982\u7387\u5206\u5e03\u6765\u9009\u62e9\u52a8\u4f5c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u80fd\u66f4\u80fd\u9f13\u52b1\u7b97\u6cd5\u63a2\u7d22\u4e0d\u540c\u7684\u52a8\u4f5c\uff0c\u5c24\u5176\u662f\u5728\u521d\u671f\u6216\u5f53\u73af\u5883\u53d8\u5316\u65f6\u3002<\/p>\n<\/li>\n
- \n
\u9700\u8981\u503c\u51fd\u6570\u903c\u8fd1\uff1aTD-Learning\u65b9\u6cd5\u901a\u5e38\u4f9d\u8d56\u4e8e\u503c\u51fd\u6570\uff08\u5982Q\u51fd\u6570\uff09\u7684\u4f30\u8ba1\u6765\u6307\u5bfc\u7b56\u7565\u7684\u9009\u62e9\u3002\u5728\u7b80\u5355\u6216\u4f4e\u7ef4\u7684\u72b6\u6001\u7a7a\u95f4\u4e2d\uff0c\u8fd9\u79cd\u4f30\u8ba1\u76f8\u5bf9\u5bb9\u6613\u3002\u7136\u800c\uff0c\u5728\u9ad8\u7ef4\u72b6\u6001\u7a7a\u95f4\u6216\u590d\u6742\u73af\u5883\u4e2d\uff08\u4f8b\u5982\uff0c\u62e5\u6709\u5927\u91cf\u72b6\u6001\u548c\u52a8\u4f5c\u7684\u6e38\u620f\uff09\uff0c\u51c6\u786e\u4f30\u8ba1\u503c\u51fd\u6570\u53d8\u5f97\u975e\u5e38\u5177\u6709\u6311\u6218\u6027\u3002<\/p>\n<\/li>\n<\/ol>\n
\u793a\u4f8b\u4ee3\u7801\uff1a\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60<\/h3>\n
\n\n- \u5b9a\u4e49\u7b56\u7565\u7f51\u7edc
\n\u7b56\u7565\u7f51\u7edc\u8f93\u51fa\u5c42\u4f7f\u7528softmax\u6fc0\u6d3b\u51fd\u6570\u751f\u6210\u52a8\u4f5c\u6982\u7387\u5206\u5e03\u3002<\/li>\n<\/ol>\n# \u5b9a\u4e49\u7b56\u7565\u7f51\u7edc\u7c7b\nclass PolicyNetwork(nn.Module):\n def __init__(self, n_states, n_actions):\n super(PolicyNetwork, self).__init__()\n self.fc1 = nn.Linear(n_states, 128)\n self.fc2 = nn.Linear(128, n_actions)\n\n def forward(self, x):\n x = F.relu(self.fc1(x))\n x = F.softmax(self.fc2(x), dim=-1) # \u4f7f\u7528softmax\u6fc0\u6d3b\u51fd\u6570\n return x\n<\/code><\/pre>\n\n- \u7b56\u7565\u68af\u5ea6\u7b97\u6cd5<\/li>\n<\/ol>\n
\u5b9e\u73b0\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\uff08\u5982REINFORCE\uff09\u66f4\u65b0\u7b56\u7565\u7f51\u7edc\u3002<\/p>\n
# \u5b9a\u4e49\u7b56\u7565\u68af\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7c7b\nclass PolicyGradient:\n def __init__(self, model, memory, cfg):\n self.n_actions = cfg['n_actions']\n self.device = torch.device(cfg['device'])\n self.gamma = cfg['gamma']\n self.policy_net = model.to(self.device)\n self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg['lr'])\n self.memory = memory\n\n # \u91c7\u6837\u52a8\u4f5c\u65b9\u6cd5\n def sample_action(self, state):\n state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(dim=0)\n probs = self.policy_net(state).cpu().detach().numpy().flatten()\n action = np.random.choice(self.n_actions, p=probs)\n return action\n\n # \u66f4\u65b0\u7b56\u7565\u7f51\u7edc\n def update(self):\n if len(self.memory) < self.memory.capacity:\n return\n state_batch, action_batch, reward_batch, _, _ = self.memory.sample(self.memory.capacity)\n state_batch = torch.tensor(np.array(state_batch), device=self.device, dtype=torch.float32)\n action_batch = torch.tensor(action_batch, device=self.device)\n reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float32)\n\n # \u8ba1\u7b97\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u6298\u6263\u56de\u62a5\n G = np.zeros_like(reward_batch)\n for t in range(len(reward_batch)):\n G_sum = 0\n discount = 1\n for k in range(t, len(reward_batch)):\n G_sum += reward_batch[k] * discount\n discount *= self.gamma\n G[t] = G_sum\n G = torch.tensor(G, dtype=torch.float32, device=self.device)\n\n self.optimizer.zero_grad()\n state_action_values = self.policy_net(state_batch)\n log_probs = torch.log(state_action_values.gather(1, action_batch.unsqueeze(1)))\n loss = -torch.sum(log_probs * G)\n loss.backward()\n self.optimizer.step()<\/code><\/pre>\n\n- \u73af\u5883\u548c\u667a\u80fd\u4f53\u914d\u7f6e<\/li>\n<\/ol>\n
\u521d\u59cb\u5316\u7b56\u7565\u68af\u5ea6\u65b9\u6cd5\u7684\u73af\u5883\u548c\u667a\u80fd\u4f53\u3002<\/p>\n
# \u73af\u5883\u548c\u667a\u80fd\u4f53\u914d\u7f6e\u51fd\u6570\ndef env_agent_config(cfg):\n env = gym.make(cfg['env_name'])\n n_states = env.observation_space.shape[0]\n n_actions = env.action_space.n\n memory = ReplayBuffer(cfg['memory_capacity'])\n agent = PolicyGradient(PolicyNetwork(n_states, n_actions), memory, cfg)\n return env, agent\n<\/code><\/pre>\n\n- \u8bad\u7ec3\u548c\u6d4b\u8bd5<\/li>\n<\/ol>\n
# \u8bad\u7ec3\u51fd\u6570\ndef train(cfg, env, agent):\n print("\u5f00\u59cb\u8bad\u7ec3\uff01")\n rewards = []\n steps = []\n for i_ep in range(cfg['train_eps']):\n ep_reward = 0\n ep_step = 0\n state, _ = env.reset()\n done = False\n while not done:\n ep_step += 1\n action = agent.sample_action(state)\n next_state, reward, done, truncated, _ = env.step(action)\n done = done or truncated\n agent.memory.push((state, action, reward, next_state, done))\n state = next_state\n ep_reward += reward\n if done:\n break\n agent.update()\n steps.append(ep_step)\n rewards.append(ep_reward)\n if (i_ep + 1) % 10 == 0:\n print(f"\u56de\u5408\uff1a{i_ep + 1}\/{cfg['train_eps']}\uff0c\u5956\u52b1\uff1a{ep_reward:.2f}")\n print("\u5b8c\u6210\u8bad\u7ec3\uff01")\n env.close()\n return {'rewards': rewards}\n\n# \u6d4b\u8bd5\u51fd\u6570\ndef test(cfg, env, agent):\n print("\u5f00\u59cb\u6d4b\u8bd5\uff01")\n rewards = []\n steps = []\n for i_ep in range(cfg['test_eps']):\n ep_reward = 0\n ep_step = 0\n state, _ = env.reset()\n done = False\n while not done and ep_step < cfg['ep_max_steps']:\n ep_step += 1\n action = agent.sample_action(state)\n next_state, reward, done, truncated, _ = env.step(action)\n done = done or truncated\n state = next_state\n ep_reward += reward\n if done:\n print(f"\u56de\u5408 {i_ep + 1} \u5728\u6b65\u6570 {ep_step} \u5b8c\u6210")\n steps.append(ep_step)\n rewards.append(ep_reward)\n print(f"\u56de\u5408\uff1a{i_ep + 1}\/{cfg['test_eps']}\uff0c\u5956\u52b1\uff1a{ep_reward:.2f}")\n print("\u5b8c\u6210\u6d4b\u8bd5")\n env.close()\n return {'rewards': rewards}\n\n# \u83b7\u53d6\u914d\u7f6e\u53c2\u6570\u51fd\u6570\ndef get_args():\n return {\n 'env_name': 'CartPole-v1',\n 'n_actions': 2,\n 'device': 'cpu',\n 'gamma': 0.99,\n 'epsilon_start': 1.0,\n 'epsilon_end': 0.01,\n 'epsilon_decay': 500,\n 'batch_size': 64,\n 'lr': 0.001,\n 'memory_capacity': 10000,\n 'train_eps': 100,\n 'test_eps': 100,\n 'ep_max_steps': 200,\n 'target_update': 10,\n 'seed': 42\n }\n\n# \u7ed8\u5236\u5956\u52b1\u66f2\u7ebf\u51fd\u6570\ndef plot_rewards(rewards, cfg, tag="train"):\n import matplotlib.pyplot as plt\n plt.plot(rewards)\n plt.title(f'{tag} Rewards')\n plt.show()\n\n# \u83b7\u53d6\u53c2\u6570\ncfg = get_args()\n# \u8bad\u7ec3\nenv, agent = env_agent_config(cfg)\nres_dic = train(cfg, env, agent)\n\n# \u7ed8\u5236\u8bad\u7ec3\u5956\u52b1\u66f2\u7ebf\nplot_rewards(res_dic['rewards'], cfg, tag="train")\n\n# \u6d4b\u8bd5\nres_dic = test(cfg, env, agent)\n\n# \u7ed8\u5236\u6d4b\u8bd5\u5956\u52b1\u66f2\u7ebf\nplot_rewards(res_dic['rewards'], cfg, tag="test")\n<\/code><\/pre>\n\u5f00\u59cb\u8bad\u7ec3\uff01\n\u56de\u5408\uff1a10\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408\uff1a20\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408\uff1a30\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408\uff1a40\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408\uff1a50\/100\uff0c\u5956\u52b1\uff1a42.00\n\u56de\u5408\uff1a60\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408\uff1a70\/100\uff0c\u5956\u52b1\uff1a41.00\n\u56de\u5408\uff1a80\/100\uff0c\u5956\u52b1\uff1a33.00\n\u56de\u5408\uff1a90\/100\uff0c\u5956\u52b1\uff1a9.00\n\u56de\u5408\uff1a100\/100\uff0c\u5956\u52b1\uff1a16.00\n\u5b8c\u6210\u8bad\u7ec3\uff01<\/code><\/pre>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926210637470.png\")
\n<\/p>\n
\u5f00\u59cb\u6d4b\u8bd5\uff01\n\u56de\u5408 1 \u5728\u6b65\u6570 31 \u5b8c\u6210\n\u56de\u5408\uff1a1\/100\uff0c\u5956\u52b1\uff1a31.00\n\u56de\u5408 2 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a2\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 3 \u5728\u6b65\u6570 23 \u5b8c\u6210\n\u56de\u5408\uff1a3\/100\uff0c\u5956\u52b1\uff1a23.00\n\u56de\u5408 4 \u5728\u6b65\u6570 20 \u5b8c\u6210\n\u56de\u5408\uff1a4\/100\uff0c\u5956\u52b1\uff1a20.00\n\u56de\u5408 5 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a5\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 6 \u5728\u6b65\u6570 21 \u5b8c\u6210\n\u56de\u5408\uff1a6\/100\uff0c\u5956\u52b1\uff1a21.00\n\u56de\u5408 7 \u5728\u6b65\u6570 29 \u5b8c\u6210\n\u56de\u5408\uff1a7\/100\uff0c\u5956\u52b1\uff1a29.00\n\u56de\u5408 8 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a8\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 9 \u5728\u6b65\u6570 25 \u5b8c\u6210\n\u56de\u5408\uff1a9\/100\uff0c\u5956\u52b1\uff1a25.00\n\u56de\u5408 10 \u5728\u6b65\u6570 47 \u5b8c\u6210\n\u56de\u5408\uff1a10\/100\uff0c\u5956\u52b1\uff1a47.00\n\u56de\u5408 11 \u5728\u6b65\u6570 27 \u5b8c\u6210\n\u56de\u5408\uff1a11\/100\uff0c\u5956\u52b1\uff1a27.00\n\u56de\u5408 12 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a12\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 13 \u5728\u6b65\u6570 20 \u5b8c\u6210\n\u56de\u5408\uff1a13\/100\uff0c\u5956\u52b1\uff1a20.00\n\u56de\u5408 14 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a14\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 15 \u5728\u6b65\u6570 32 \u5b8c\u6210\n\u56de\u5408\uff1a15\/100\uff0c\u5956\u52b1\uff1a32.00\n\u56de\u5408 16 \u5728\u6b65\u6570 17 \u5b8c\u6210\n\u56de\u5408\uff1a16\/100\uff0c\u5956\u52b1\uff1a17.00\n\u56de\u5408 17 \u5728\u6b65\u6570 15 \u5b8c\u6210\n\u56de\u5408\uff1a17\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408 18 \u5728\u6b65\u6570 27 \u5b8c\u6210\n\u56de\u5408\uff1a18\/100\uff0c\u5956\u52b1\uff1a27.00\n\u56de\u5408 19 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a19\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 20 \u5728\u6b65\u6570 26 \u5b8c\u6210\n\u56de\u5408\uff1a20\/100\uff0c\u5956\u52b1\uff1a26.00\n\u56de\u5408 21 \u5728\u6b65\u6570 26 \u5b8c\u6210\n\u56de\u5408\uff1a21\/100\uff0c\u5956\u52b1\uff1a26.00\n\u56de\u5408 22 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a22\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 23 \u5728\u6b65\u6570 16 \u5b8c\u6210\n\u56de\u5408\uff1a23\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408 24 \u5728\u6b65\u6570 15 \u5b8c\u6210\n\u56de\u5408\uff1a24\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408 25 \u5728\u6b65\u6570 43 \u5b8c\u6210\n\u56de\u5408\uff1a25\/100\uff0c\u5956\u52b1\uff1a43.00\n\u56de\u5408 26 \u5728\u6b65\u6570 10 \u5b8c\u6210\n\u56de\u5408\uff1a26\/100\uff0c\u5956\u52b1\uff1a10.00\n\u56de\u5408 27 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a27\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 28 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a28\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 29 \u5728\u6b65\u6570 15 \u5b8c\u6210\n\u56de\u5408\uff1a29\/100\uff0c\u5956\u52b1\uff1a15.00\n\u56de\u5408 30 \u5728\u6b65\u6570 34 \u5b8c\u6210\n\u56de\u5408\uff1a30\/100\uff0c\u5956\u52b1\uff1a34.00\n\u56de\u5408 31 \u5728\u6b65\u6570 16 \u5b8c\u6210\n\u56de\u5408\uff1a31\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408 32 \u5728\u6b65\u6570 30 \u5b8c\u6210\n\u56de\u5408\uff1a32\/100\uff0c\u5956\u52b1\uff1a30.00\n\u56de\u5408 33 \u5728\u6b65\u6570 13 \u5b8c\u6210\n\u56de\u5408\uff1a33\/100\uff0c\u5956\u52b1\uff1a13.00\n\u56de\u5408 34 \u5728\u6b65\u6570 16 \u5b8c\u6210\n\u56de\u5408\uff1a34\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408 35 \u5728\u6b65\u6570 13 \u5b8c\u6210\n\u56de\u5408\uff1a35\/100\uff0c\u5956\u52b1\uff1a13.00\n\u56de\u5408 36 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a36\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 37 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a37\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 38 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a38\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 39 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a39\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 40 \u5728\u6b65\u6570 41 \u5b8c\u6210\n\u56de\u5408\uff1a40\/100\uff0c\u5956\u52b1\uff1a41.00\n\u56de\u5408 41 \u5728\u6b65\u6570 24 \u5b8c\u6210\n\u56de\u5408\uff1a41\/100\uff0c\u5956\u52b1\uff1a24.00\n\u56de\u5408 42 \u5728\u6b65\u6570 26 \u5b8c\u6210\n\u56de\u5408\uff1a42\/100\uff0c\u5956\u52b1\uff1a26.00\n\u56de\u5408 43 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a43\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 44 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a44\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 45 \u5728\u6b65\u6570 16 \u5b8c\u6210\n\u56de\u5408\uff1a45\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408 46 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a46\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 47 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a47\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 48 \u5728\u6b65\u6570 13 \u5b8c\u6210\n\u56de\u5408\uff1a48\/100\uff0c\u5956\u52b1\uff1a13.00\n\u56de\u5408 49 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a49\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 50 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a50\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 51 \u5728\u6b65\u6570 21 \u5b8c\u6210\n\u56de\u5408\uff1a51\/100\uff0c\u5956\u52b1\uff1a21.00\n\u56de\u5408 52 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a52\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 53 \u5728\u6b65\u6570 13 \u5b8c\u6210\n\u56de\u5408\uff1a53\/100\uff0c\u5956\u52b1\uff1a13.00\n\u56de\u5408 54 \u5728\u6b65\u6570 17 \u5b8c\u6210\n\u56de\u5408\uff1a54\/100\uff0c\u5956\u52b1\uff1a17.00\n\u56de\u5408 55 \u5728\u6b65\u6570 54 \u5b8c\u6210\n\u56de\u5408\uff1a55\/100\uff0c\u5956\u52b1\uff1a54.00\n\u56de\u5408 56 \u5728\u6b65\u6570 20 \u5b8c\u6210\n\u56de\u5408\uff1a56\/100\uff0c\u5956\u52b1\uff1a20.00\n\u56de\u5408 57 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a57\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 58 \u5728\u6b65\u6570 26 \u5b8c\u6210\n\u56de\u5408\uff1a58\/100\uff0c\u5956\u52b1\uff1a26.00\n\u56de\u5408 59 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a59\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 60 \u5728\u6b65\u6570 10 \u5b8c\u6210\n\u56de\u5408\uff1a60\/100\uff0c\u5956\u52b1\uff1a10.00\n\u56de\u5408 61 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a61\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 62 \u5728\u6b65\u6570 30 \u5b8c\u6210\n\u56de\u5408\uff1a62\/100\uff0c\u5956\u52b1\uff1a30.00\n\u56de\u5408 63 \u5728\u6b65\u6570 10 \u5b8c\u6210\n\u56de\u5408\uff1a63\/100\uff0c\u5956\u52b1\uff1a10.00\n\u56de\u5408 64 \u5728\u6b65\u6570 22 \u5b8c\u6210\n\u56de\u5408\uff1a64\/100\uff0c\u5956\u52b1\uff1a22.00\n\u56de\u5408 65 \u5728\u6b65\u6570 24 \u5b8c\u6210\n\u56de\u5408\uff1a65\/100\uff0c\u5956\u52b1\uff1a24.00\n\u56de\u5408 66 \u5728\u6b65\u6570 52 \u5b8c\u6210\n\u56de\u5408\uff1a66\/100\uff0c\u5956\u52b1\uff1a52.00\n\u56de\u5408 67 \u5728\u6b65\u6570 51 \u5b8c\u6210\n\u56de\u5408\uff1a67\/100\uff0c\u5956\u52b1\uff1a51.00\n\u56de\u5408 68 \u5728\u6b65\u6570 31 \u5b8c\u6210\n\u56de\u5408\uff1a68\/100\uff0c\u5956\u52b1\uff1a31.00\n\u56de\u5408 69 \u5728\u6b65\u6570 39 \u5b8c\u6210\n\u56de\u5408\uff1a69\/100\uff0c\u5956\u52b1\uff1a39.00\n\u56de\u5408 70 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a70\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 71 \u5728\u6b65\u6570 17 \u5b8c\u6210\n\u56de\u5408\uff1a71\/100\uff0c\u5956\u52b1\uff1a17.00\n\u56de\u5408 72 \u5728\u6b65\u6570 36 \u5b8c\u6210\n\u56de\u5408\uff1a72\/100\uff0c\u5956\u52b1\uff1a36.00\n\u56de\u5408 73 \u5728\u6b65\u6570 21 \u5b8c\u6210\n\u56de\u5408\uff1a73\/100\uff0c\u5956\u52b1\uff1a21.00\n\u56de\u5408 74 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a74\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 75 \u5728\u6b65\u6570 28 \u5b8c\u6210\n\u56de\u5408\uff1a75\/100\uff0c\u5956\u52b1\uff1a28.00\n\u56de\u5408 76 \u5728\u6b65\u6570 28 \u5b8c\u6210\n\u56de\u5408\uff1a76\/100\uff0c\u5956\u52b1\uff1a28.00\n\u56de\u5408 77 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a77\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 78 \u5728\u6b65\u6570 37 \u5b8c\u6210\n\u56de\u5408\uff1a78\/100\uff0c\u5956\u52b1\uff1a37.00\n\u56de\u5408 79 \u5728\u6b65\u6570 12 \u5b8c\u6210\n\u56de\u5408\uff1a79\/100\uff0c\u5956\u52b1\uff1a12.00\n\u56de\u5408 80 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a80\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 81 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a81\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 82 \u5728\u6b65\u6570 14 \u5b8c\u6210\n\u56de\u5408\uff1a82\/100\uff0c\u5956\u52b1\uff1a14.00\n\u56de\u5408 83 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a83\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 84 \u5728\u6b65\u6570 37 \u5b8c\u6210\n\u56de\u5408\uff1a84\/100\uff0c\u5956\u52b1\uff1a37.00\n\u56de\u5408 85 \u5728\u6b65\u6570 16 \u5b8c\u6210\n\u56de\u5408\uff1a85\/100\uff0c\u5956\u52b1\uff1a16.00\n\u56de\u5408 86 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a86\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 87 \u5728\u6b65\u6570 17 \u5b8c\u6210\n\u56de\u5408\uff1a87\/100\uff0c\u5956\u52b1\uff1a17.00\n\u56de\u5408 88 \u5728\u6b65\u6570 32 \u5b8c\u6210\n\u56de\u5408\uff1a88\/100\uff0c\u5956\u52b1\uff1a32.00\n\u56de\u5408 89 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a89\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 90 \u5728\u6b65\u6570 45 \u5b8c\u6210\n\u56de\u5408\uff1a90\/100\uff0c\u5956\u52b1\uff1a45.00\n\u56de\u5408 91 \u5728\u6b65\u6570 35 \u5b8c\u6210\n\u56de\u5408\uff1a91\/100\uff0c\u5956\u52b1\uff1a35.00\n\u56de\u5408 92 \u5728\u6b65\u6570 36 \u5b8c\u6210\n\u56de\u5408\uff1a92\/100\uff0c\u5956\u52b1\uff1a36.00\n\u56de\u5408 93 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a93\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 94 \u5728\u6b65\u6570 33 \u5b8c\u6210\n\u56de\u5408\uff1a94\/100\uff0c\u5956\u52b1\uff1a33.00\n\u56de\u5408 95 \u5728\u6b65\u6570 18 \u5b8c\u6210\n\u56de\u5408\uff1a95\/100\uff0c\u5956\u52b1\uff1a18.00\n\u56de\u5408 96 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a96\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 97 \u5728\u6b65\u6570 32 \u5b8c\u6210\n\u56de\u5408\uff1a97\/100\uff0c\u5956\u52b1\uff1a32.00\n\u56de\u5408 98 \u5728\u6b65\u6570 11 \u5b8c\u6210\n\u56de\u5408\uff1a98\/100\uff0c\u5956\u52b1\uff1a11.00\n\u56de\u5408 99 \u5728\u6b65\u6570 19 \u5b8c\u6210\n\u56de\u5408\uff1a99\/100\uff0c\u5956\u52b1\uff1a19.00\n\u56de\u5408 100 \u5728\u6b65\u6570 20 \u5b8c\u6210\n\u56de\u5408\uff1a100\/100\uff0c\u5956\u52b1\uff1a20.00\n\u5b8c\u6210\u6d4b\u8bd5<\/code><\/pre>\n\n ![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926210700498.png\")
\n<\/p>\n
![](\"https:\/\/img.icons8.com\/?size=100&id=aaIXU3iIXlkB&format=png&color=5C7CFA\")
\u6f14\u5458\u8bc4\u8bba\u5bb6\u6a21\u578b\uff08Actor-Critic model\uff09<\/h2>\n
\n
\u6f14\u5458-\u8bc4\u8bba\u5bb6\uff08Actor-Critic\uff09\u6a21\u578b\u662f\u4e00\u79cd\u7ed3\u5408\u4e86\u57fa\u4e8e\u503c\u7684\u65b9\u6cd5\u548c\u57fa\u4e8e\u7b56\u7565\u7684\u65b9\u6cd5\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u3002 \u8fd9\u4e2a\u6a21\u578b\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u7b56\u7565\u51b3\u7b56\uff08\u6f14\u5458\uff09\u548c\u503c\u51fd\u6570\u4f30\u8ba1\uff08\u8bc4\u8bba\u5bb6\uff09\u7684\u4f18\u70b9\u7ed3\u5408\u8d77\u6765\uff0c\u4ee5\u671f\u8fbe\u5230\u66f4\u597d\u7684\u5b66\u4e60\u6548\u7387\u548c\u7b56\u7565\u6027\u80fd\u3002<\/p>\n
\u6f14\u5458\u90e8\u5206\u8d1f\u8d23\u57fa\u4e8e\u5f53\u524d\u7b56\u7565\u9009\u62e9\u52a8\u4f5c<\/strong>\u3002<\/p>\n
![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926164437847.png\")
Deep Learning<\/h1>\n![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926164522677.png\")
\n<\/p>\n![](\"https:\/\/img.icons8.com\/bubbles\/50\/000000\/checklist.png\")
Agenda<\/h3>\n![](\"https:\/\/img.icons8.com\/cute-clipart\/64\/000000\/alarm.png\")
\u5f3a\u5316\u5b66\u4e60\u7684\u57fa\u672c\u539f\u7406<\/h2>\n![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926164609304.png\")
\n<\/p>\n![](\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240926171139868.png\")
\n<\/p>\n