引入优势函数A,优势函数A = 状态动作价值函数Q - 状态价值函数V。
在同一状态下,所有动作的优势值为零。因为,所有的动作的状态动作价值的期望就是状态价值。
实现代码:
import random import gym import torch import numpy as np from matplotlib import pyplot as plt from IPython import display env = gym.make("Pendulum-v0") # 智能体状态 state = env.reset() # 动作空间 actions = env.action_space print(state, actions) # 打印游戏 # plt.imshow(env.render(mode='rgb_array')) # plt.show() """重新定义策略价值网络Q, 比DQN性能更优""" class VAnet(torch.nn.Module): def __init__(self): super().__init__() self.fc = torch.nn.Sequential(torch.nn.Linear(3, 128), torch.nn.ReLU()) self.fc_A = torch.nn.Linear(128, 11) self.fc_V = torch.nn.Linear(128, 1) def forward(self, x): A = self.fc_A(self.fc(x)) V = self.fc_V(self.fc(x)) A_mean = A.mean(dim=1).reshape(-1, 1) A = A - A_mean # Q值由A和V求和得到 Q = A + V return Q # 定义动作模型(策略网络) model = VAnet() # 经验网络,评估一个动作的分数(目标网络) next_model = VAnet() # model的参数赋予next_model next_model.load_state_dict(model.state_dict()) # 得到一个动作 def get_action(state): """state: agent所处的状态。由于是连续动作,做离散化操作""" # 走神经网络NN,得到分值最大的那个动作。转为tensor数据 state = torch.FloatTensor(state).reshape(1, 3) action = model(state).argmax().item() if random.random() < 0.01: action = random.choice(range(11)) # 离散动作连续化 action_continuous = action action_continuous /= 10 action_continuous *= 4 action_continuous -= 2 return action, action_continuous # 数据池 datas = [] def update_data(): """加入新的N条数据,删除最老的M条数据""" count = len(datas) while len(datas) - count < 200: # 一直追加数据,尽可能多的获取环境状态 state = env.reset() done = False while not done: # 由初始状态开始得到一个动作 action, action_continuous = get_action(state) next_state, reward, done, _ = env.step([action_continuous]) datas.append((state, action, reward, next_state, done)) # 更新状态 state = next_state # 此时新数据集中比原来多了大约200条样本,如果超过了最大容量,删除最开始数据 update_count = len(datas) - count while len(datas) > 5000: datas.pop(0) return update_count # 从数据池中采样 def get_sample(): # batch size = 64, 数据类型转换为Tensor samples = random.sample(datas, 64) state = torch.FloatTensor([i[0] for i in samples]).reshape(-1, 3) action = torch.LongTensor([i[1] for i in samples]).reshape(-1, 1) reward = torch.FloatTensor([i[2] for i in samples]).reshape(-1, 1) next_state = torch.FloatTensor([i[3] for i in samples]).reshape(-1, 3) done = torch.LongTensor([i[4] for i in samples]).reshape(-1, 1) return state, action, reward, next_state, done # 获取动作价值 def get_value(state, action): """根据网络输出找到对应动作的得分,使用策略网络""" action_value = model(state) action_value = action_value.gather(dim=1, index=action) return action_value # 获取学习目标值 def get_target(next_state, reward, done): """使用next_state和reward计算真实得分。对价值的估计,使用目标网络""" with torch.no_grad(): target = next_model(next_state) target = target.max(dim=1)[0].reshape(-1, 1) target *= (1 - done) # 游戏结束的状态,没有奖励 target = reward + target * 0.98 return target # 一局游戏得分测试 def test(): reward_sum = 0 state = env.reset() done = False while not done: _, action_continuous = get_action(state) next_state, reward, done, _ = env.step([action_continuous]) reward_sum += reward state = next_state return reward_sum def train(): model.train() optimizer = torch.optim.Adam(model.parameters(), lr=2e-3) loss_fn = torch.nn.MSELoss() for epoch in range(600): # 更新一批数据 update_counter = update_data() # 更新过数据后,学习N次 for i in range(200): state, action, reward, next_state, done = get_sample() # 计算value和target value = get_value(state, action) target = get_target(next_state, reward, done) # 参数更新 loss = loss_fn(value, target) optimizer.zero_grad() loss.backward() optimizer.step() """周期性更新目标网络""" if (i + 1) % 10 == 0: next_model.load_state_dict(model.state_dict()) if epoch % 50 == 0: test_score = sum([test() for i in range(50)]) / 50 print(epoch, len(datas), update_counter, test_score)
标签:06,torch,next,state,done,action,DQN,reward,Dueling From: https://www.cnblogs.com/demo-deng/p/16889897.html