Dueling DQN or DDQN extends the concept of a fixed target or fixed Q target and extends that to include a new concept called advantage. Advantage is a concept where we determine what additional value or advantage we may get by taking other actions. Ideally, we want to calculate advantage so that it includes all the other actions. We can do this with computational graphs by separating the layers into a calculation of state value and another that calculates the advantage from all the permutations of state and action.
This construction can be seen in the following diagram:
DDQN visualized in detail
The preceding ...
