I appreciate the exciting sudden influx of AI games we have been lacking. I hope to try out some play against it and maybe learn something too.

Hey hang. I just finished a master's research project in a few state-of-the-art rl models. I am trying to enter the field so I am building a portfolio. I would love to talk to you, learn how you coded it, and improve the model.

啥时候能干time？

Thanks for working on this I think it's awesome! I am curious about two things: How does this differ from the bot that Scarlett helped program in terms of logic? How much compute power is needed to run this, could I run my own AI league at home and it would just take longer to get batches done or does it need multiple servers worth of compute?

There is a UKposts channel named Laughingamez that did a lot of Alphastar content. He would be interested in replays.

This is the most relatable comment I've come across

that would be realy great!

Yeah, great idea!

I don't know if he would be up for the challenge

he work professionally. You gotta pay him for this.

I'd love this

Humans spam for two reasons AI don't have. One is that humans don't react instantly to the map and jumping your camera around gives you higher chance of randomly stumbling into something just out the minimap. The other... is just keeping your hands warm.

@@stamythezombie Also keeps you more focused on the game

@@stamythezombie The Carpal Tunnel Syndrome just isn't worth it.

Also ego and average apm?? Lol

yea because these AI´s allways cheat on the higher difficulty settings, but an AI like this is a fascinating conceot, it could very well be the future of gaming

Unfortunately running aplhastar is really expensive.

When I see videos like this I think the AI overlord that will end up turning us all into batteries will come from a video game.

@@SpielkindFR Dont worry, humans are shit for batteries, even potatoes are better.

That was the first Alphastar. The ones they released for ladder were ridiculously nerfed and had severe APM caps.

@@Figgy20000 Yup, but it was still doing micro not possible with a mouse and keyboard, that was the point. The stalker micro with the capped APM still isn't possible movement for a pro. In fact, that was one of the telltale signs when people were hunting for that AI in the ladder. Compared to this Zerg AI that does things like a human pro.

@@N0noy1989 Incorrect, DM adjusted the parameters so that AS did not exceed what a top GM can do both on ladder and at the Blizz event. iirc even some players below GM won a game.

@@dannygjk They limited its average and burst APM, but it still had times with much more EFFECTIVE APM since real humans use autoclickers and spam clicking a lot, which inflated those caps on the AI, and which the AI could use for pure effective actions rather than the high proportion of redundant actions humans do.

@@goatmeal5241 DM nerfed AS even more than that. Look it up.

Imagine what it will do to humans in RL

if you train him to do it, he will ;)

他俩毕竟是爸爸跟儿子😂

MTCA!

thats funny because alphastar is literally the most useless uninteresting application of machine learning i have ever encountered in my career as an AI engineer. glad to see it at least did some good despite being a bigger exercise in futility than coding a chess computer in the 21st century to prove computers are better at math than humans. alphastar won all its games with unrealistic apm and micro accuracy and inhuman response times. no amount of gfuel can get a human to play that fast, even if apm is limited to what humans can realistically achieve its not a fair game, thus the computer will probably win.

@@TheSuperappelflap It exposed me to machine learning, the thought that an AI could make decisions based on data it had collected and refer back to it in real time absolutely captivated me

@@TheSuperappelflap Surprising to me that an "AI engineer" would say some of the things you said. For example, "...to prove computers are better at math than humans.". There is a lot more to neural net systems than just doing math better (faster?) than humans.

@@dannygjk neural networks are just engines for doing a lot of linear algebra in parallell, which is also how the human brain works. the difference is that human brains are very complex with billions of nodes performing different tasks that are all interconnected, and neural networks are designed to perform a single task in isolation, therefore they are much better at performing that one task at very high speeds. designing a neural network that can beat humans at one video game by performing a couple tasks very well very quickly is not a challenge in 2022. its been proven many times over the past 40 years that humans are no competition for computer engines at games, from chess computers to alphago. alphastar for sc2 was just a PR project for google, nothing more. it was a waste of time and money that they could have invested in building a neural network that can do something useful.

@@TheSuperappelflap Your explanation, (which I was already aware of), is missing an explanation of why human biases is not a factor compared to the old school techniques.

If he prioritizes lurkers over the infestor texh he’d almost certainly win but I think he lacked scouting info. You really need to do everything possible with zerg to know what your opponent is doing. He definitely had that game with like 6-10 lurkers and maybe ten more roaches to tank. He also didn’t know about the push until it was too late if the infestors were ready with more than 1 fungal it could have done a lot of damage to the roaches. I think the lurker tech is more reliable and even just 5 lurkers on that ramp backed by roaches would do gobtons of damage. Being down a carapace upgrade was also painful but I get money was tight since he was trying to texh his way to victory. He honestly lost the moment he wasn’t ready for the mutas. You can tell he’s not a zerg player because even diamond players would have those spores down. Spore queen buys time for your hydras it doesn’t beat mutas. It’s a lack of understanding the tech flow in the matchup. Lurker hydra then viper is sort of end game so you need to optimize towards those techs which help swing the game either getting you to late game or crushing the roaches, mutas, and/or ling bane player outright. Alternatively he could have gone for his own mutas and used the home court advantage + carapace upgrade to get the edge while teching into bane speed and crushing eco with LBM. I think hydra tech is generally more stable and it feeds into lurkers which are the Trump card for Zerg ground. Again it’s obvious he doesn’t play zerg because it was all very sloppy. On a side note I think a roach wall off build would crush this AI. It’s predictably going bane nest + lings pretty heavily and I’m not sure the AI would even know how to stop this build (it’s not easy if you just commit to lings blindly they can expand/drone up or go for timing.

Losing your economy to 6 roaches is not a master plan. I also played against the AI and won by abusing it a bit more. It definitely very much lacks an understanding of its position in the game and when and where to attack.

I think it was waiting for a bank so it could instantly replace the mutas when they die quickly in a straight up battle and remax for a 300 supply push. it actually worked tbf. It was a "mistake" however to let eco die to the roaches when it could have sent 3 mutalisk back to kill them.

I feel the AI suffers a bit from the same fate though. It looks like every model is just optimized for one strategy, and will just go for that strategy every time. That means you could have a bunch of different models and a random one being selected every time to get some sort of game variety, but it still misses the true reactivity that a human opponent can provide. I don't think for example that a model would ever decide to switch things up and go a different strategic direction based on what the opponent is doing.

Technically this is how the AI learns to play in the first place, by playing a near endless amount of games against itself and adjusting its behavior based on what does and doesn't work. The version of the AI you are seeing here is basically the champion of a million bot tournament.

The SC2 siege tank already does this to some extent automatically (not perfectly though) in comparison to the SC1 tank, it makes sure to spread fire to not overkill as much.

Check the description! "iAsonu changed the names of the models to Sc II characters: ModelA-rl_model, modelB-Abuthur, modelC-Brakk, modelD-Dehaka"

@@FFAs should have read more haha

So Harstem said no limitations. Seems really unfair, not even just the APM, also the lack of window view restriction and as a result global action.

@@MagnificentXXBastard I dont think Starcraft has any API for outside AI so any 3rd party AI needs to use human UI to control it and "see". Unlimited APM sounds like a bad idea since it can just jump around with screen to have unlimited view and control. That being said, there seems to be some kind of limitation since it keeps itself in normal APM numbers.

@@belisarian6429 I am pretty sure that that was a thing they introduced to alphastar after some criticism that it actually had to use & move the window to play now instead of global actions.

@@MagnificentXXBastard Really? Ok just quick googled it and it seems 3rd party AIs can indeed use API instead of human interface. That is sad, as that means that those AIs are not as good as I anticipated as this removes difficulties and limitations that come with human interface and need to be artificially added, like for instance camera movement.

Most likely AI simply ignores that chat and does not even consider it as variable that is in its inputs. So it will likely just get discarded as unnecessary noise or nonsense and never even reach decision making neurons.

There was a big AI project to train it against a bunch of old Atari games (simple mechanics, most games have a displayed score). When taught how to play Tetris and that higher score = good, losing = bad, whenever the AI was going to lose, it would just pause the game. Can't die on the pause screen.

There was a big AI project to train it against a bunch of old Atari games (simple mechanics, most games have a displayed score). When taught how to play Tetris and that higher score = good, losing = bad, whenever the AI was going to lose, it would just pause the game. Can't die on the pause screen.

I think the main reason we dont see this behavior is, that the AIs cant surrender, just like chatting. Its not an essential feature for playing the game, so the Devs leave it out. And after that, your points become important. IF the AIs had the option to leave, why would they? They have "infinite" time, cant DC unless their training machines crash, and would actively lose the slim chance to win. The only reasonable surrender option I could think of would be: "No attacking units left and no resources to produce more". As a hard-coded behavior and not learned.

There is, time is time, and playing out a game where you have effectively zero chance of winning is time not spent in a game where you can win

@@protoss5758 For humans yes, but AIs or Computer Programs in general arent like us. These AIs only "exist" during a game or in training. The other time they are just a static file on a hard drive. So for them there is no time between matches. Even matches might not be the same. AlphaStar for example plays on an optimized version for computer meaning no UI, way faster game tick rate etc. And we have yet reached a point in AI development where adding time constraints makes sense and does not just add more potential sources of error.

I don't think AS has been publicly active since Blizzcon.

Link in the description

There are more different possibilities in sc2 though and also limited information which actually benefits the humans because in Chess the engine can easily calculate every possible consequence of every possible move, but in sc2 as a real time strategy, the possibilities are closer to infinite, especially as a game of imperfect information which creates a use for actual prediction, bluffs, misdirection and stuff that isn't really a factor for a chess bot because in Chess the AI has immediate perfect information of every possibility and every move which happens. If the SC2 bot is limited by the fog of war & not maphacking, that factor alone makes it much more complicated.

Figuring out how to teach AI to have theory of mind isn't a big problem, and whichever AI is given theory of mind definitely wouldn't be this StarCraft AI with huge tactical and strategic abilities