What is a DQN policy?

In the module stable_baselines.deepq.policies, you have a lot of policies for DQN. This is all very nice, but Q-learning (and DQN) is an off-policy and value-based method, which means that we should be learning a state-action value function with a behaviour policy (like epsilon-ungreedy) that is derived from that value function (which, of course, does not need parameters, apart from the greek letter), not a policy, so what the heck do these policies represent? Are they actually the state-action value function Q(s, a) (in the case of the DQN, a neural network), or what? That’s the only thing in DQN that has parameters anyway, so I guess that all those “policies” there are actually neural networks that represent state-action value functions Q(s, a). Right? The documentation is as useless as it could ever be:

Policy object that implements a DQN policy

Yes, I did not look at the code yet.

[araffin]

Note: dqn only has a q network, so it would be: net_arch=[]

[Miffyli]

@sheila-janota Yup!