def learn(self, obs, action, reward, next_obs, terminal): """ 使用DQN算法更新self.model的value网络 """ # 从target_model中获取 max Q' 的值,用于计算target_Q next_pred_value = self.target_model.value(next_obs) best_v = layers.reduce_max(next_pred_value, dim=1) best_v.stop_gradient = True # 阻止梯度传递 terminal = layers.cast(terminal, dtype='float32') target = reward + (1.0 - terminal) * self.gamma * best_v pred_value = self.model.value(obs) # 获取Q预测值 # 将action转onehot向量,比如:3 => [0,0,0,1,0],独热编码有好处 action_onehot = layers.one_hot(action, self.act_dim) action_onehot = layers.cast(action_onehot, dtype='float32') # 下面一行是逐元素相乘,拿到action对应的 Q(s,a) # 比如:pred_value = [[2.3, 5.7, 1.2, 3.9, 1.4]], action_onehot = [[0,0,0,1,0]] # ==> pred_action_value = [[3.9]] pred_action_value = layers.reduce_sum(layers.elementwise_mul( action_onehot, pred_value), dim=1) # 计算 Q(s,a) 与 target_Q的均方差,得到loss cost = layers.square_error_cost(pred_action_value, target) cost = layers.reduce_mean(cost) optimizer = fluid.optimizer.Adam(learning_rate=self.lr) # 使用Adam优化器 optimizer.minimize(cost) return cost
def learn(self, obs, action, reward, next_obs, terminal): next_pred_value = self.target_model.value(next_obs) best_v = layers.reduce_max(next_pred_value, dim=-1) best_v.stop_gradient = True terminal = layers.cast(terminal, dtype="float32") target = reward + (1.0 - terminal) * self.gamma * best_v pred_value = self.model.value(obs) action_onehot = layers.one_hot(action, self.act_dim) action_onehot = layers.cast(action_onehot, dtype="float32") pred_action_value = layers.reduce_sum(layers.elementwise_mul( pred_value, action_onehot), dim=-1) cost = layers.square_error_cost(target, pred_action_value) cost = layers.reduce_mean(cost) optimizer = fluid.optimizer.Adam(learning_rate=self.lr) optimizer.minimize(cost) return cost
def learn(self, obs, action, reward, next_obs, terminal): ''' :param obs: St :param action: At :param reward: Rt+1 :param next_obs: St+1 :param terminal: done, True代表episode结束 :return: 损失函数的值 ''' # 通过目标网络计算得到target_Q的值 target_Q_tensor = self.target_model.value(next_obs) # 计算St+1对应的价值向量 max_Q = layers.reduce_max(target_Q_tensor, dim=1) # 获取每行的最大值,按dim=1收缩 max_Q.stop_gradient = True # 停止梯度更新 # 由于terminal不是标量,所以不能直接用判断 terminal = layers.cast(terminal, dtype="float32") target_Q = reward + (1.0 - terminal) * self.gamma * max_Q # 通过主网络计算得到perdict_Q的值 predict_Q_tensor = self.model.value(obs) # 将action转成one-hot向量,并将每一位都变成浮点数 action_onehot = layers.one_hot(action, self.act_dim) action = layers.cast(action_onehot, dtype="float32") # 进行elementwise计算并降低张量阶数 # 比如 predict_Q_tensor = [[2.3, 5.7, 1.2, 3.9, 1.4], action_onehot=[[0, 0, 0, 1, 0] # [2.1, 3.7, 4.5, 6.7, 7.1]] [0, 1, 0, 0, 0]] # 那么elementwise乘法运算后的结果是 [[0, 0, 0, 3.9, 0] # [0, 3.7, 0, 0, 0]] # 再进行dim=1的reduce_sum操作后的结果是 [3.9, 3.7] predict_Q = layers.reduce_sum(layers.elementwise_mul( action_onehot, predict_Q_tensor), dim=1) # 得到这个batch每条数据的损失函数值的平均值 cost = layers.square_error_cost(predict_Q, target_Q) cost = layers.reduce_mean(cost) # 申明优化器(使用Adam优化器) optimizer = fluid.optimizer.Adam(learning_rate=self.lr) optimizer.minimize(cost) # 指定优化目标 return cost
def get_greedy_prob(scores_padded, mask_padded): s = scores_padded - (mask_padded * (-1) + 1) * self.BIG_VALUE max_value = layers.reduce_max(s, dim=1, keep_dim=True) greedy_prob = layers.cast(s >= max_value, 'float32') return greedy_prob