Hacker News new | past | comments | ask | show | jobs | submit login
Super Smash Bros. Melee: An “Untouchable” Agent (arxiv.org)
64 points by jasonjmcghee on Dec 14, 2017 | hide | past | favorite | 26 comments



In the discussion they mention:

> The major limitation of our project was the opposing player. Of course, we would have preferred training against human players who play Melee professionally.

An alternative training method could be through the use of the 20XX hack pack [0], a mod of Super Smash Bros Melee that adds, among many useful features, an improved AI. Here's a video of the AI from release 4.05 in action: https://www.youtube.com/watch?v=rzl5NXV5bHQ

The built-in SSBM AI is very noticeably different from competitive play, and the 20XX AI makes significant steps toward more human-like competitive play. It would be interesting to see the results if trained on these AI.

[0] https://smashboards.com/threads/the-20xx-melee-training-hack...


Why not just put it on netplay? Faster Melee makes strides towards latency improvements but I think a little latency and instability would actually be better to teach the AI fundamentals like spacing and zoning, rather than just beating opponents in terms of timing.


> The built-in SSBM AI is very noticeably different from competitive play

And not necessarily that clever. Here are examples of the CPU losing to a player who does nothing:

https://www.youtube.com/watch?v=CwDngBhmQXY


This is not an example of dumb AI. cpus will lose to a non moving Luigi in more than one smash game. It's intentional behavior.


Can you prove that it's intentional? It looks more like a very specific positioning of the character relative to the map and a lot of trial and error that is causing the cpu to kill itself


Sure, look at this video. This is the latest game in the series.

https://youtu.be/SFBnhaXql24


Holy crap after digging deeper this is actually just a meme I thought was an actual thing for like the last 7 or 8 years. Whoops


Are you sure that it is intentional?


This guy is memeing.


https://github.com/altf4/SmashBot

Here is an AI that beat a ton of pro melee players. Video here https://youtu.be/o1bfQWy8o08


"Our modern ML AI lasted almost a minute against 15 y.o. hand-coded AI, 74.5% of the time!"


If you just wanted the videos: https://goo.gl/x67ioE.


If you're on a phone, here's an HTML version of the paper: https://www.arxiv-vanity.com/papers/1712.03280/


Thanks, the video link in the pdf doesn't work, but the one here does: https://drive.google.com/drive/folders/0B9SX5ANPyMzTNWgtMzdq...


Mildly confused:

The paper claims there are only 5 used actions, but I also observed wavedashing.

Also as an aside this is yet another in a series of not as cool as I'd hoped melee ML projects :(


I was pretty impressed by the Falcon AI earlier this year. He actually used modern reinforcement learning to train from 0 instead of the crappy built in AI. Also, his AI was limited to near-human reaction time (10 frames) which led to a more human-looking agent.

At the end of the day these papers are from master's students. I think they're both impressive in that context.

https://arxiv.org/pdf/1702.06230.pdf


Interestingly the reason they chose falcon was that he has no projectile attacks. I also remember hearing that it could only play against other falcons, and only on battlefield. But it could have been improved since then


> Falcon AI

which one is that, I hadn't seen it (and its not the one you linked).


"The agent has access to five actions to choose from: do nothing, left dodge, right dodge, standing dodge, and Fox’s shine (essentially a shield that pushes opponents away)."

Possibly classified as left/right dodge?


It requires a shine, jump, and then a left/right dodge executed in a specific time frame, which makes it seem like it was programmed in. My guess is that they just wanted to focus on the deep learning and not get bogged down in Melee mechanics/terminology.


Left/right dodges are the rolls. Wavedashes are a more complex maneuver.


Wave dashing is a combination of actions, not a single action.


It requires jumping though, which isn't listed as an action


Well its a series of actions not a combination.


Google Drive says the videos have exceeded some limit and can no longer be played. Which, to me, suggests that maybe Google Drive is not the best place to host videos.


> It learned to avoid the toughest AI built into the game for a full minute 74.6% of the time

No results against humans, though.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: