I am grateful for your clean and well-documented open-source Agent!
I noticed a possible mistake in the Rainbow-Agent README.md. In the README, Dueling Networks is defined as "Dividing the neural net stream into two branches, an action stream and a value stream. Both of them combined form the Q-action values." However, this definition is implemented under "adversarial" in the code. Additionally, the provided paper defines Double Q-learning as "reducing overestimations by decomposing the max operation in the target into action selection and action evaluation," which seems to refer to separating the Target model from the main model in agent.py
Hello @ClementPerroud,
I am grateful for your clean and well-documented open-source Agent!
I noticed a possible mistake in the Rainbow-Agent README.md. In the README, Dueling Networks is defined as "Dividing the neural net stream into two branches, an action stream and a value stream. Both of them combined form the Q-action values." However, this definition is implemented under "adversarial" in the code. Additionally, the provided paper defines Double Q-learning as "reducing overestimations by decomposing the max operation in the target into action selection and action evaluation," which seems to refer to separating the Target model from the main model in agent.py