Excellent!

In the timescale of the next week, I'm only hoping to get a very specific letter from someone at the company that says literally "If the proposal submitted by Dr. [insert the full name of the Principal Investigator] entitled [insert the proposal title] is selected for funding by the NSF, it is my intent to collaborate and/or commit resources as detailed in the Project Description or the Facilities, Equipment or Other Resources section of the proposal." I'm still writing the proposal now (deciding whether to mention Itch by name if I can get a letter, or just "public archive and marketplaces" otherwise). The only resources I actually need from Itch are rough agreement that you are okay with the effort to use your public data and willingness to answer questions about whether I'm breaking stuff -- super minimal. They letter mostly just shows that I've checked in with you folks and that my research plan won't come crumbling down the moment I try to start it up.

Anything beyond that, such as sending one of my graduate students there as an intern to replicate our prototype systems on top of your production stuff, can be saved for far far off in the future.

One of our first projects in this area used screenshots collected while playing back speedrun input files: https://adamsmith.as/papers/KEG-18_paper_4.pdf (tested on Super Mario World and Super Metroid). It showed that you could search by screenshots and find specific results within a given game, but it begged the question: where do you get the speedruns from?

In another project, we extracted screenshots and audio transcripts from YouTube Let's Play videos: https://adamsmith.as/papers/GAMNLP-19-3.pdf We could find the "horses" in Last of Us and the "selfie" in Life is Strange, but still, can a machine do the playing instead of a human?

Of course a machine can -- and even really dumb strategies can sometimes do well. We tried out some simple strategies on games spanning from Atari 2600 to Nintendo 64: https://adamsmith.as/papers/KEG_2019_paper_9.pdf

But, like, can the algorithm cover serious ground on some timescale that compares with human gameplay? We looked at that in Super Mario World and Legend of Zelda: https://adamsmith.as/papers/PID5953685.pdf It turns out that both humans and machines loose steam as you let them explore for longer. Interleaving human and machine exploration lead to the most amount of ground covered for a given amount of time invested.

Do these things just press random buttons or does it learn from experience? Our first algorithms just pressed random buttons, yes, but the new ones (tested on in-development Unity games that integrate our machine playtesting script) do learn from experience and improve over time: https://adamsmith.as/papers/MonsterCarlo2.pdf They can bootstrap from human gameplay demonstration or from their own random flailing.

We're in the business of creating all this stuff and just giving it away for free. My lab brings in money by proposing interesting new work to be done (to be paid for by taxpayers) rather than selling the technology we developed last year. I'm scratching on the door of Itch as part of a proposal to get more money (from the US National Science Foundation) to do the next steps in these projects. It's tricky to get funding for research around games, but we're trying.