The moving target problem is due to continuously updating the network during training, which also modifies the target values. Nevertheless, the neural network has to update itself in order to provide the best possible state-action values. The solution that's employed in DQNs is to use two neural networks. One is called the online network, which is constantly updated, while the other is called the target network, which is updated only every N iterations (with N usually being between 1,000 and 10,000). The online network is used to interact with the environment while the target network is used to predict the target values. In this way, for N iterations, the target values that are produced by the target network remain fixed, preventing the propagation of instabilities and decreasing the risk of divergence. A potential disadvantage is that the target network is an old version of the online network. Nonetheless, in practice, the advantages greatly outweigh the disadvantages and the stability of the algorithm will improve significantly.