Hi Arka, as I mentioned in the text — you copy the first network into the target network every once in a while. In a regular DQN learning, the first predictions that are made come from pure random weights, yet you use them for training. That’s the same thing here, just on longer periods.