Added Policy Gradient Tutorial

[shreyas-kowshik]

*Description
Implementation of Vanilla Monte Carlo Policy Gradients on the CartPole-v0 environment added as a tutorial.

*Tests
Run the script

[tejank10]

opt = ADAM(params(policy),η) could be
opt = ADAM(η) in accordance with the new optimizer API

[tejank10]

Broadcasting is not required for a variable element.

[tejank10]

WRT the new Optimizer API, this will become update!(opt, params(model))

[shreyas-kowshik]

@tejank10

update!(opt,params(policy))

does not find a matching candidate.
Tried using

grads = Tracker.gradient(() -> loss(state,act,G_t), params(policy))

for p in params(policy)
  update!(opt, p, grads[p])
end

but even this throws up errors.

[DhairyaLGandhi]

Please add a Project.toml and Manifest.toml as well so it easier to standardize the environment.

[shreyas-kowshik]

@dhairyagandhi96 Added the files

[MikeInnes]

Can we just use normal headings for these rather than the extra formatting / html tags?

[MikeInnes]

@tejank10 are you happy with the changes made here, or is there more to do?

[shreyas-kowshik]

@MikeInnes Thanks for the reply. I apologize for making a few errors before. The DCGAN code should not have been included in this PR. There is a separate PR for that. I have corrected it by removing the GAN code. The changes you mentioned for the GAN part will be updated in the respective PR.

[shreyas-kowshik]

@tejank10 I have made the changes requested. Sorry for having delayed this for so long. I got into other work and did not fix the errors that were coming up. The changes have been completed now. I have also added functions to normalize the discounted rewards which would aid in training the network.

[MikeInnes]

It will also need to be in its own folder, and have a simple README. Otherwise this is looking good I think, but it'd be good to hear from @tejank10.

[shreyas-kowshik]

@MikeInnes Sorry for the delayed response. I have made the changes. Is the README sufficient for now or is there something more to be added?