Reinforcement Learning with TensorFlow

Dueling DQN

In case of Dueling DQN, the Q value has been modified as the summation of the value function of the state and advantage function of the action. The value function V(s) quantifies the usefulness or goodness of being in state s and the advantage function A(a) quantifies the advantage of action a over other possible actions. Therefore,

Dueling DQN has separate networks to compute the value and advantage functions and then combine them back to fetch the value for the Q-function. The reason behind decoupling the computation of value and advantage is that the agent doesn't have to take care of the unnecessary value function computations ...