Implemented DQN and Double DQN with GPU support

[sidhantnagpal]

This adds DQN and Double DQN using Reinforce.jl with GPU support. Key highlights (tested on Colab with Julia 1.1):

• Batching (after sampling from replay memory) for Q-learning
• Demo recording (mp4) (ffmpeg is required for the same)
• Model saving/loading (bson)
• Model testing (with minimum exploration)

[MikeInnes]

Seems like a nice feature set! Can you add a readme with walkthrough instructions and a project/manifest?

[sidhantnagpal]

Thanks for having a look. What instructions are exactly expected in the readme? (For reference, I have the notebooks at: https://colab.research.google.com/github/sidhantnagpal/model-zoo/blob/nbs/) Also, how can I access the project/manifest toml files on Colab?

[MikeInnes]

You can look at other models in the zoo for inspiration. Just some background and an idea of what the model is expected to show is helpful. I don't know much about colab myself.

[sidhantnagpal]

Turns out the files can be moved from the Colab filesystem to the mounted drive easily. The toml files were accessible under ~/.julia/environments/v1.1/. I have included them.

I have also added the problem description and walkthrough in the readme.