Investigate Training a Mega Man GAN using Mega Man Maker Levels

[schrum2]

I want to see how well a GAN can be trained to produce Mega Man levels based on Mega Man Maker data. I think @cappsb already trained some GANs on certain Mega Man Maker levels. They are in the src\main\python\GAN\MegaManGAN directory and start with the prefix MMLV. I'm hoping @cappsb can give us some guidance on what exactly is here.

Maxx's first job here is to see if any of those GANs work with the current code. Try loading the MegaManGANLevelBreederTask but change the GAN parameters to be these MMLV GANs. If those GANs don't work for some reason, then we shouldn't keep them in the repository. If they do work, then I'm curious to know if the levels look interesting.

Ultimately, we want to see if we can train on lots of MMLV levels rather than just the small set of Mega Man 1 levels from VGLC.

[MBatt1]

With issues fixed in #657, the MMLV GANs appear to work

[schrum2]

We may still want to make some MMLV GANs trained on more levels. Hopefully, @cappsb will give us some kind of indication regarding what levels were trained on. It would also be good to do some experiments to see if the MMLV levels are qualitatively different from the VGLC GANs. Ultimately, the long-term goal is to use a large set made up of VGLC and MMLV levels to see how well the GAN does.

Side note: @MBatt1 : Look though the closed issues at some point to see if there are any notes about how the MMLV GANs were trained. You could also look at commit messages associated with those files

[cappsb]

@schrum2 @MBatt1 From what I remember, I only trained levels based on the "Examples" from mmlv, with a couple of omissions because of levels with a lot of special blocks that we do not account for.

[schrum2]

I still want to train on more MMLV levels eventually, but I think this issue can rest for now.

[schrum2]

@MuellMark Will need to get Mega Man Maker up and running first to make sense of this. It's a big task, but could be really cool to do.