I was reading the article Deep Reinforcement Learning: Playing CartPole through Asynchronous Advantage Actor-Critic (A3C) with tf.keras and eager execution. From my understanding, we copy the weights of the "worker model" to the "global model".
But won't this erase the work of other workers?
For example, if $W_1$ (worker 1) updates the weight to the global model and it will start again processing the next episode, and, after that, $W_2$ (worker 2) will copy weights to the global model. That will erase the work of $W_1$.
Am I correct? If yes, then how is this problem handled? If no, then what is right?
