July 2018
Intermediate to advanced
254 pages
8h 12m
English
In Double Q-Learning, there are two action-value/Q functions. Let's call them Q1 and Q2. The idea in double Q-learning is to decouple action selection from the value estimation. That is, when we want to update Q1, we select the best action according to Q1, but use Q2 to find the value of the selected action. Similarly, when Q2 is being updated, we select the action based on Q2, but use Q1 to determine the value of the selected action. In practice, we can use the main Q network
as Q1 and the slow-moving target network
as Q2, ...
Read now
Unlock full access