Optimizing Our Way Through Metroid

Yeah the company sounds interesting. I wish the main page had clearer info about what it does. There’s a lot of text but I want the simple, “here’s the little bit of example code to get going.”

Haven’t tried Castlevania II, but here’s the first one: https://antithesis.com/blog/castlevania/

  Castlevania... [so] called because it is a Metroidvania game set in a Castle.

> I would love to see how it handles Castlevania II.

I assume you're thinking specifically of using the red crystal to spawn a tornado: https://youtu.be/Mx9PwRIK9Io

I thought the same thing. They are quite verbose in explaining themselves but I found their docs to be useful.

https://antithesis.com/docs/

Curious - what OS? (I work at Wistia!)

The longer you run it, the cleaner the run gets. But Metroid is a very compute-intensive game to fuzz, and we were already nearing the limits of what BigQuery could do for us with that run.

There have definitely been some applications of this sort of thing to speedrunning -- though far less sophisticated than the approach here, and usually only testing against a very small subset of the game. I've heard of some of this kind of work being done before on e.g. SM64.

I've also done something along these lines myself in Super Metroid. Mother Brain's neck moves in a conceptually simple but very chaotic pattern influenced by Samus's vertical movement, and there's a cutscene during the fight where the positioning of her neck can make a difference of about 7 seconds. The TAS fight used complicated movement to manipulate her neck position developed through much trial-and-error, while the best known human-viable manips were several seconds slower.

I wrote a program to search the state space for optimal movement patterns, and working with some speedrunners we were able to come up with a new human-viable manipulation that matched the previous TAS fight, as well as a new TAS manipulation that saved an additional 41 frames.

https://youtu.be/7SHD9L_Jx5Q

https://github.com/NobodyNada/mbsim

Yes, this is a really fun idea and something that we want to do. Though these days we’re setting our sights higher than Nintendo…

A funny story though: a regular conference gimmick we have is “Man vs. Machine” where we have attendees race our fuzzer to the end of Mario level 1-1. We did this at the final year of Strange Loop, and the fuzzer was winning handily until not one, not two, but three different professional speedrunners walked by and destroyed us.

Adding 1 CSS rule gets rid of the slowness:

  background-color: black !important;

I'm not sure which specific one is to blame, but there is a lot of transparency in various colors, both foreground and background.

It did remind me of that. Tom7 is a treasure

It’s not quite a fair comparison, since an RL agent is trying to learn a policy that wins fair and square, while a fuzzer is able to take back moves. But if you’re working in a domain (like anything that can be simulated) where “time travel” is possible, you’d have to be crazy not to use it!