First Chess, then Go, and now Super Smash Bros. Melee?! Are the AIs finally getting the better of us at everything?

Not quite, but close enough!

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have created a Super Smash Bros. Melee AI superior to the regular "target practice" you can already fight in the game. And it’s good enough to hold its own against some global top 100 players!

Super Smash Bros. Melee was chosen because, according to the paper's abstract, its "environment has complex dynamics and partial observability, making it challenging for human and machine alike. The multiplayer aspect poses an additional challenge"

The team, led by Vlad Firoiu, trained a neural network model to play the game by feeding it gameplay data and incentivizing play that resulted in the computer’s victory. Rather than watching the screen and learning patterns, it is more like an in-game computer player that’s learned everything from scratch.

“It uses a combination of human techniques and some odd ones too – both of which benefit from faster-than-human reflexes,” wrote Firoiu in an email to TechCrunch. “It is sometimes very conservative, being unwilling to attack until it sees there’s a opening. Other times it goes for risky off-stage acrobatics that it turns into quick kills.”

However, it doesn't fare well with projectiles and it has a secret weakness:

“If the opponent crouches in the corner for a long period of time, it freaks out and eventually suicides,” Firiou wrote. (“This should be a warning against releasing agents trained in simulation into the real world,” he added)

Below is a video of the AI playing against several top 100 global players, against whom it won more than it lost!

​

While not a reason to fret for your gamer skills (yet?), the research demonstrates the interesting way in which AIs can cope in new environments. Moreover, game devs could make use of such neural networks to provide a more serious challenge to competitive and professional players to practice their skills.

SOURCE

 

[Jak27]

More importantly, they should share this AI with the rest of the world, that way Melee players all over can improve even further in the comfort of their homes. I know I would love be able to practice against something like this on a regular basis.

 

[Vorde]

Sooo much ignorance and cancer in this thread. I thought more of you gbatemp community, smh

 

[Veho]

Moreover Artificial intelligence is not intelligence at all it is a program.

If you think so you have zero understanding of artificial intelligence.

 

[Foxi4]

They've barely scratched the surface in understanding intelligence so it's pointless to be trying to duplicate it. Contrary to popular belief your brain is not a computer
https://aeon.co/essays/your-brain-does-not-process-information-and-it-is-not-a-computer

I had psychology in uni too, but this isn't amateur hour. The human brain stores conceptualized information rather than words, that much is correct, but it doesn't invalidate the previous statement. You are born with certain instinctual reflexes, which goes against your theory instead of supporting it. You react to stimuli because you're genetically programmed to - you can't outsmart fear or pain, it's embedded programming.

 

[FAST6191]

The AI doesn't look that much good. Maybe he's making the best choices, but it just can't do things like the pros to..
For example, wavedashing (or if we're talking about capt. Falcon, so dash-dancing and SHFFL'd nairs to approach)
Also it uses side smash with Falcon too much.. it might be good sometimes, but not that much.
Alsp near the end you can see the AI just doing strange sh*t.. It actually used up-b after up air which means he maybe programed to recover if you don't land after aerial-attacks
There's a lot to improve, but I can see it going well even in the near future.

A few years back there was a nice AI developed to play backgammon. When the dust had settled a lot of the received wisdom on the best modes of play lay in tatters on the floor (not the one I am thinking of but something along the same lines http://www.cs.cornell.edu/boom/2001sp/Tsinteris/gammon.htm ). Are you sure you are not heading down a similar path?

Artificial intelligence will never be as good as natural intelligence.

---
Moreover Artificial intelligence is not intelligence at all it is a program.

The computing power of the brain is finite, not to mention seriously odd in a lot of ways*, and you can always stitch more silicon together, humans already do not design computers and have computers design computers and a lot of what were considered complex problems are solved by computers. 3d and pattern recognition is still a problem but there are massive strides being made into it. To that end why will AI not match humans at some point?

Also the matter of intelligence or not is a rather philosophical one. I prefer to take a more practical approach of what can it do? To that end intelligence or not seems rather immaterial.

 

[DarkCoffe64]

Bet Luigi can still win by doing absolutely nothing.

 

[Imacaredformy2ds]

Bet Luigi can still win by doing absolutely nothing.

He actually can

 

[anhminh]

Now we only need to make android with this kind of AI for military purpose.

 

[Hayleia]

Melee's following is beyond dedicated. Imagine fusing the collective egos' of PC gamers, Arch Linux, and every single fanboy, into one community. It still won't reach the level's of Melee's community.

The reason for this is simple. There was never a sequel to Arch (or PC or whatever you want) that basically killed all what Arch (or PC or whatever you want) was liked for. Now of course, devs fixed in sequels what they didn't do on purpose in Melee, I'm not saying they should have done Melee 2 or whatever. Just saying that Melee fans are not just fan of it, they're also and most of all salty that the sequels are far from it, that's why the Melee community looks like this and not the Arch/PC/whatever community.

 

[Gamemaster1379]

lmao everyone with his point of view.. If you don't that much understand the meta-game, you'll find yourself asking why are they spamming shines.
It excites me more than sf5 and any other fighting game

As a former competitive player, it's mentalities like yours that kill the interest from a community standpoint.

Meta should be, but isn't about what is "best". It might start with a fundamental "best", at least at the time, but it generally devolves, not evolves. Top players become stuck in their ways, and often don't have to change or adapt because of the reputation that precedes them. Many players of lesser skill blindly follow and uphold a "meta" standpoint. If they play, they often become monkeys who emulate actions without understanding as to why they do what they do.

A proper meta should evolve based on perception. Often time it evolves around the current meta expectation and the best way is to manipulate those expectations and take advantage of them--eventually becoming the meta--and thus being overriden.

In my time as a competitive player, I specialized in psychology, perceiving what other players thought they should or would do--and manipulating that. I often took teams out of their comfort zones, making them have to try and adapt on the fly--and often times, they fall short. It's more powerful and interesting to be adaptive and have a changing, fluid metagame, rather than having a set "standpoint" and that being the be-all-to-end-all without any consideration of "non-meta" strategies.

 

[Hayleia]

As a former competitive player, it's mentalities like yours that kill the interest from a community standpoint.

Meta should be, but isn't about what is "best". It might start with a fundamental "best", at least at the time, but it generally devolves, not evolves. Top players become stuck in their ways, and often don't have to change or adapt because of the reputation that precedes them. Many players of lesser skill blindly follow and uphold a "meta" standpoint. If they play, they often become monkeys who emulate actions without understanding as to why they do what they do.

A proper meta should evolve based on perception. Often time it evolves around the current meta expectation and the best way is to manipulate those expectations and take advantage of them--eventually becoming the meta--and thus being overriden.

In my time as a competitive player, I specialized in psychology, perceiving what other players thought they should or would do--and manipulating that. I often took teams out of their comfort zones, making them have to try and adapt on the fly--and often times, they fall short. It's more powerful and interesting to be adaptive and have a changing, fluid metagame, rather than having a set "standpoint" and that being the be-all-to-end-all without any consideration of "non-meta" strategies.

That's why I don't get how a reinforcement learning AI could beat "the best" at Melee. Yeah, it could beat the "monkeys" you talk about, but all the ones who spend time studying their opponent will still be able to wreck it. Maybe they'll have to get 4-stock 10 times before understanding what works, but they'll figure it out and be able to 4-stock it back one day.

Example: have you noticed how the AI, when lying on the ground (missed tech) almost always (I only noticed 1 time where it didn't do it (at 5:09 [edit: it did it once too at the beginning, and the two times it rolled towards the edge when it was close to it already, so that might actually be another pattern its opponent can take advantage of...]) after it struck me that it almost always did it) gets up with an attack? Not rolls left not rolls right, not just stands up, almost always attacks. Which ultimately leads to what happens at 5:22, the AI is on the ground, Slox knows how it is going to get up so he just has to jump in the meantime to land a knee once the AI is vulnerable. See, just sudying that predictable opponent would make it beatable by any pro who actually reads and not just spam techniques learnt on wikihow.

Even worse. Who are you going to train it against? M2K? Then it will have no clue what Bizzarro Flame does and will get wrecked by him even though he would never win against M2K. Or you'll train it against Bizz? Then it will get wrecked by people who face the direction they run towards instead of moonwalking all the time. Or against Mango? Then he'll stop dashdancing and just walk up slowly and down smash... And I'm not even mentioning the fact these players don't main the same character.

However, I'm not saying this AI is shit, I'd love to play against it instead of that stupid AI we have in Melee which has a reaction time of 0 (so Falco's lasers are useless against it) but 0 skill (you can kill a lvl 9 Falcon just by using your shield on the edge, you don't even need to attack). So yeah, I welcome this project, but it shouldn't expect to be unbeatable.

 

[Haloman800]

The grey captain falcon is the AI, right?

 

[Gamemaster1379]

That's why I don't get how a reinforcement learning AI could beat "the best" at Melee. Yeah, it could beat the "monkeys" you talk about, but all the ones who spend time studying their opponent will still be able to wreck it. Maybe they'll have to get 4-stock 10 times before understanding what works, but they'll figure it out and be able to 4-stock it back one day.

Example: have you noticed how the AI, when lying on the ground (missed tech) almost always (I only noticed 1 time where it didn't do it (at 5:09 [edit: it did it once too at the beginning, and the two times it rolled towards the edge when it was close to it already, so that might actually be another pattern its opponent can take advantage of...]) after it struck me that it almost always did it) gets up with an attack? Not rolls left not rolls right, not just stands up, almost always attacks. Which ultimately leads to what happens at 5:22, the AI is on the ground, Slox knows how it is going to get up so he just has to jump in the meantime to land a knee once the AI is vulnerable. See, just sudying that predictable opponent would make it beatable by any pro who actually reads and not just spam techniques learnt on wikihow.

Even worse. Who are you going to train it against? M2K? Then it will have no clue what Bizzarro Flame does and will get wrecked by him even though he would never win against M2K. Or you'll train it against Bizz? Then it will get wrecked by people who face the direction they run towards instead of moonwalking all the time. Or against Mango? Then he'll stop dashdancing and just walk up slowly and down smash... And I'm not even mentioning the fact these players don't main the same character.

However, I'm not saying this AI is shit, I'd love to play against it instead of that stupid AI we have in Melee which has a reaction time of 0 (so Falco's lasers are useless against it) but 0 skill (you can kill a lvl 9 Falcon just by using your shield on the edge, you don't even need to attack). So yeah, I welcome this project, but it shouldn't expect to be unbeatable.

I find AI inherently interesting. They don't pre-suppose metagame, they go off observation, and putting it against top players--it inherently capitalizes anti-meta as these players all play and revolve around it.

As it progresses, it will eventually counter any playstyle it plays against long enough, reasonably speaking (unless a bug was found that is exploited). But, fundamentally speaking, it advocates diversity of playstyle and evolution of gameplay.