Average DQN implementation

[rbauld]

This is my attempt to implement averaged DQN, based on this paper: https://arxiv.org/pdf/1611.01929.pdf

By default with avg_dqn=1 It should just be regular DQN/DDQN. When avg_dqn> 1 it will average over previous models. So you gain some stability at the cost of some computation time.

In any case, let me know what you think, or what tests should be done. I have already ran the agent tests and the cart pole example, but I think more extensive testing may be needed.

Seemingly the testing branch is 6 months behind, so I guess everyone is just merging to master?

[RaphaelMeudec]

Hi! Thanks for your contribution, tests are failing since a few days ago, I'll dig into it and then look at your implementation!

[rbauld]

I took a look at some of the errors on the travis CI build. They don't really seem related to any of the changes I have made.

Seems other pull requests have had a similar issue.

In any-case, the changes were not too drastic, but I thought they were worthwhile making, since I have found some benefit when using average DQN in particularly noisy environments.