I don’t know much about the Alphastar project. I gathered a few information here and there from what I could find on the internet and I find this topic very intriguing.
I watched it play (and win) against some definitly top players of Zerg and Protoss but unfortunantly I didn’t find games vs top Terrans.
As far as I heard Alphastar is being considered an actual artificial intelligence (AI!!!) by google. A Software that is intentionally handycaped by it’s creators to use limited apm (Action per Minute) to match human apm and yet capable to learn and adept to the degree that it can keep up with GM level players.
I would like to know what happened to Alphastar? And why it was removed. Was it even removed?
I would like to see a version of Alphastar being implimented as a “bot” that you can play against in melee mode. I mean what other games are out there where you can best yourself against an actual AI? That’s pretty cool isn’t it?
Starcraft 2 is the only game that one I know of that went this far with experimental AI. I think it would help the games longlivity for a future to come if a version of Alphastar was added as a “bot” with Googles permission to custom games.
Still it is hard for me to believe that it is an actual AI… It may act intelligent, but it’s not a free mind is it? I might get a bit philosophical here… There is so much I’d like to know about Alphastar.
As I understood, it got shut down because further supporting it costs a lot of resources. I guess it’s not just a random PC running in a basement, but something a lot more. So just adding it to the game is not that easy as it may sound.
I strongly suspect they used a genetic algorithm for training the bots. So basically they generated a million or so randomized bots and had them compete on a virtual ladder. Then they discard the bottom 1-5% and fill those slots in with bots that are mixtures of the the top 1-5%. They repeat this many tens of thousands of times and eventually have a bot that can do a single build on a single map, with absolutely zero capacity to learn and change itself.
So the AI system is a “two part” system - the AI that trains the bots, and the actual bots that play the actual games. The bots would have to be constantly retrained for meta shifts or new maps, which costs $$$ because you need hundreds of super computers going 24/7.
mayb as smart as protons or zergs alphastar never show playing terran becuse terran 2 hard 4 dum computer not as smart as terrans 2 learn 2 play terran
TLO who worked with the Alphastar team said in an interview, or pylon show. That Google accomplished what they set out to do with alphastar and are very happy with the results, so they are not working on that anymore. Apparently this and alphazero (chess) among other projects are aimed towards gathering more info to develop AIs.
I dont like that idea tho. What we do in battle net is violent in its core nature. Its about survival of fittest. However human understand its a game but ai doesnt.
You are vastly overestimating AIs, they do not have any sort of consciousness or feelings, as of now they just do as they are told through algorithms. Too much sci fi my friend
That’s not new. The Perimet’r System designed several decades ago is tasked with unleashing the full Power of the Nuclear potential (enough to end life in Earth several times) even after all the human component is killed or incapacitated.
It’s the Samson Doctrine.
No research group is ever going to say that they weren’t happy with the results. Some critical inspection of their methodology finds very large flaws. They published an initial iteration that had many advantages over humans and they tried to fix a few of them. It had inhuman micro capabilities for example. In the “final” iteration it still had significantly more average APM than humans do, it could click on the very edge of the screen (something that would cause your camera to move if you were human), etc.
The biggest limitation is that in SC2 the meta changes. The meta is the composition of all strategies that are being used, how often, and in what circumstances. The meta changes on its own, because players are always modifying their play to counter the most common strategies that they face. The meta also changes dramatically when a new strategy is found, or when the map pool is changed. New strategies are found all the time, since they are highly map dependent. Terrans will find a new place to put a siege tank for a particular timing that just completely rewrites the meta.
Alphastar was 100% shielded from the meta. It’s obvious why - they would have to spend $$$ to retrain it constantly as the meta shifted since the individual bots have zero capacity to learn / adapt on their own. Being in GM or a Pro means being able to be one step ahead of the meta so SC2 is very much a game of learning faster than your opponents and having a better memory of all the classifications of opponents and each of their recent playstyles. Alphastar never played enough on the ladder to contribute to the meta, and if it did it would almost certainly be maybe diamond league (maximum). Players would have an opportunity to figure out how it plays, and to craft strategies that prey on the weaknesses that it has in its play.
That’s the thing about SC2 - there is no “right” strategy - there are only strategies that win more than they lose and that depends on the meta. So they were basically training this AI against thousands of human replays, giving it tons of data on the human meta, but humans were given basically zero opportunity to ever do the same. It can’t be understated just how massive of an advantage this is for Alphastar. It’s like if you were going into a game knowing exactly how your opponent plays, having crafted the perfect strategy to beat him, meanwhile he has no clue who you are or how you play.
Alphastar was a big step forward but it’s -nowhere- near what they portray it to be and you can’t really call it AI. It’s a very complicated statistical prediction system, and that’s it. There is no “learning” component to it. It’s literally about flinging crap at a wall and seeing what sticks and what doesn’t. It maps a set of inputs to a set of outputs using very complicated matrix math and that’s it.