the necessity of copying current_q?

[Bennett561]

I dont quite see the necessity of copying current_q in func train (dqn_agent).. since you just need the shape of it, right, why not use np.zeros_like...?

    def train(self, batch):
        """
        Trains the underlying network with a batch of gameplay experiences to
        help it better predict the Q values.
        :param batch: a batch of gameplay experiences
        :return: training loss
        """
        state_batch, next_state_batch, action_batch, reward_batch, done_batch \
            = batch
        current_q = self.q_net(state_batch).numpy()
        target_q = np.copy(current_q)
        next_q = self.target_q_net(next_state_batch).numpy()
        max_next_q = np.amax(next_q, axis=1)
        for i in range(state_batch.shape[0]):
            target_q_val = reward_batch[i]
            if not done_batch[i]:
                target_q_val += 0.95 * max_next_q[i]
            target_q[i][action_batch[i]] = target_q_val
        training_history = self.q_net.fit(x=state_batch, y=target_q, verbose=0)
        loss = training_history.history['loss']
        return loss

[SebastianKotstein]

According to my understanding the step of calculating current Q-values matrix and copying them into target Q-Values matrix is required since the following for loop updates only those elements of the target Q-Values matrix whose indices match the chosen actions. Therefore, all other elements of the target Q-Values matrix must be equal to the values of the current Q-values matrix, otherwise, for instance, if target Q-Values would be initialized with zeros or random values, the optimizer would try to fit the model against these zero or random values.