[Armag] diarouga wrote:This is just lame. The AI just makes full stalkers and wins because perfect macro in sc2 means that you don't lose a single unit.
Yes. This just seemed like the shooter variant of 'just click heads' and win. Strategically it seemed okay, I think it basically understands the game at a ~masters level. That's still impressive, because it basically discovered all of this on its own, but it doesn't seem like this AI managed to break strategic barriers here. They need to tone down the apm, maybe even half it, and try to restrict its awareness/multitasking.
Yea definitely, the AI has 270 APM, but it's actually 270 EPM, which is more than anyone but 2-3 players in the world, and with no missclic, it can see the whole map, and perfect APM spenging. It should be limited to like 150-200 EPM lol.
MAybe even 100 epm or less. If literally every click is accurate, that's unlike any human. Humans will often click a few times for actions and spend a ton of apm actually navigating the map. This AI is basically godlike mechanically with it's top down view and something thats probably equivalent to say a 700 apm human player.
How didn't it win easily? He was in control from the moment he started pushing Mana's base. All the stalker lost were clearly taken into account as part of the strategy. He would have changed unit perhaps only vs a perfect definding AI but even then he would still keep the advantage of being the offender. Basically in AI vs AI the other AI wouldn't do Mana's strat but probably it would be a mirror strat since going stalker bullying is just better. How isn't this about strategy? Obviously there is a huge gap in mechanics that means the AI could win with a variety of strats, but that doesn't mean the AI isn't trying its best picking the best strat for the situation it faces.
Well if it did this in 10 days of training, don't you guys think in 6 months it will be by fat the absolute best player in the world, both mechanically and strategically?
Ye, I was skeptical about it, but considering this was only 2weeks of training (even tho it's months of developing) I think less than 1 year is enough to create an AI that wins consistently vs humans even with super capped APM and all ohter human limitations. But anyway with this kind of AIs it's all about testing it vs other AIs. Humans are not even remotely close to win a single game without mechanical capping.
Garja wrote:How it didn't win easily? He was in control from the moment he started pushing Mana base. All the stalker lost were clearly taken into account as part of the strategy. He would have changed unit perhaps only a perfect definding AI but even then he would still keep the advantage of being the offender. Basically in AI vs AI the other AI wouldn't do this strat but probably it would be a mirror strat since going stalker bullying is just better. How isn't this about strategy? Obviously there is a huge gap in mechanics that means the AI could win with a variety of strats, but that doesn't mean the AI isn't trying its best picking the best strat for the situation it faces.
No, it wouldn't go full stalker. As they explain in the stream, they have different versions of the AI, which all have a different play style and use different build orders. They of course made the AIs play against each other and there was no AI that dominated. It means that all the strategies which were used in these games counter some others and are countered by some others, so this stalker mass isn't the perfect strategy.
ovi12 wrote:Well if it did this in 10 days of training, don't you guys think in 6 months it will be by fat the absolute best player in the world, both mechanically and strategically?
I'm not sure they said that no one dominated. They said it was hard to say which one is stronger. I'm not sure the top5 they picked actually played a lot vs each other and they tracked their win rates. Btw top5 were the less exploitable, they said.
Also I didn't mean that two AIs would do stalker mass. I said mirror strat because it's a mirror MU. More generally they would just do the nash equilibrium strat, which in case of a mirror it's same strat. Being incomplete information game it means they would act as in a static game so they have to pick best strat from the beginning and only adapt from there.
The average APM is limited to 270, which in practice is 270 EPM. Considering that less than 10 players in the game have more than 270 EPM and it has zoom hack (which means that camera moves don't count in that APM), it's like 330 EPM, which is more than anyone but Serral, the AI is super fast.
Also it's just average APM, which means that the AI can has 800 APM during a fight, which is lame.
[Armag] diarouga wrote:Now without zoom hack, AlphaStar got molested by mana. And mana is probably like top50.
Yes, that proves our points right. The zoom hack and apm were overtuned.
The AI looked like a super strong player though that did some things really well.
Yea, AlphaStar played really well, like a middle GM probably, but not on pro level. As they said in the stream, it would be great to see the AI playing every races to finally prove that protoss is op
I thought the move to attack with stalkers in the natural while going in the main with the oracle was a very 'human' move that seems based on human limitations. This AI might still be a bit overtuned though, as it still has about 270 EPM which is insane. Also apperantly playing effectively for 200 years didn't teach it that you can make a phoenix against a warp prism. So strategically speaking it just seems like it can approach a smart human being, but is still not as competent.
momuuu wrote:I thought the move to attack with stalkers in the natural while going in the main with the oracle was a very 'human' move that seems based on human limitations. This AI might still be a bit overtuned though, as it still has about 270 EPM which is insane. Also apperantly playing effectively for 200 years didn't teach it that you can make a phoenix against a warp prism. So strategically speaking it just seems like it can approach a smart human being, but is still not as competent.
Yea, as I said earlier, probably master's level strategically. Also it ignores the counter system which doesn't make sense.
Honestly, it just seemed like a strategically mediocre player now that I consider it. I think I have more understanding of the game but I can't really make things work because I don't have 270 epm. I mean the midgame oracles were just straight up stupid and I think anyone in gold/platinum understands that full stalker vs immortals doesn't work.
It's not like it ignores the counter system, it just knows better that some units don't actually counter others well enough with OP micro involved. A bit like going sepoy or ashi vs skirms in AOE3, attacking from more than one angle.
That AI is very impressive. If it can get to this level you can imagine it can go further no problem.
Some things are odd but still work. For example it seems to value presaturating before expanding. But to always be well marcoed and to make the combat decisions it is making is very impressive. Obviously it’s through testing 200 years of possible ways to interpret these scenarios. But things like the ramp block are so fascinating because there’s so many ways to build a base and it’s not entirely obvious to a brute force situation why you would ever reach outside of building in base.
It also showcases how you have more options if you have better mechanics. It can win situations a worse mechanical player could not. Mechanics and strategy in humans are linked and in the game they are also linked. It’s not possible to separate them entirely because the level of depth to your game expands as your mechanics open up new avenues.
_H2O wrote:That AI is very impressive. If it can get to this level you can imagine it can go further no problem.
Some things are odd but still work. For example it seems to value presaturating before expanding. But to always be well marcoed and to make the combat decisions it is making is very impressive. Obviously it’s through testing 200 years of possible ways to interpret these scenarios. But things like the ramp block are so fascinating because there’s so many ways to build a base and it’s not entirely obvious to a brute force situation why you would ever reach outside of building in base.
It also showcases how you have more options if you have better mechanics. It can win situations a worse mechanical player could not. Mechanics and strategy in humans are linked and in the game they are also linked. It’s not possible to separate them entirely because the level of depth to your game expands as your mechanics open up new avenues.
_H2O wrote:That AI is very impressive. If it can get to this level you can imagine it can go further no problem.
Some things are odd but still work. For example it seems to value presaturating before expanding. But to always be well marcoed and to make the combat decisions it is making is very impressive. Obviously it’s through testing 200 years of possible ways to interpret these scenarios. But things like the ramp block are so fascinating because there’s so many ways to build a base and it’s not entirely obvious to a brute force situation why you would ever reach outside of building in base.
It also showcases how you have more options if you have better mechanics. It can win situations a worse mechanical player could not. Mechanics and strategy in humans are linked and in the game they are also linked. It’s not possible to separate them entirely because the level of depth to your game expands as your mechanics open up new avenues.
_H2O wrote:That AI is very impressive. If it can get to this level you can imagine it can go further no problem.
Some things are odd but still work. For example it seems to value presaturating before expanding. But to always be well marcoed and to make the combat decisions it is making is very impressive. Obviously it’s through testing 200 years of possible ways to interpret these scenarios. But things like the ramp block are so fascinating because there’s so many ways to build a base and it’s not entirely obvious to a brute force situation why you would ever reach outside of building in base.
It also showcases how you have more options if you have better mechanics. It can win situations a worse mechanical player could not. Mechanics and strategy in humans are linked and in the game they are also linked. It’s not possible to separate them entirely because the level of depth to your game expands as your mechanics open up new avenues.
Didnt the bot mana played train for even more than 200 ingame years? I wonder if its capped in its potential. I mean, 200 years of playing and the bots havent figured out that you can distract someone by dropping immortals with a warp prism. Talk about a slow learner haha.
I badly want to see them let this bot grind much more and then let it play with half the apm.
_H2O wrote:That AI is very impressive. If it can get to this level you can imagine it can go further no problem.
Some things are odd but still work. For example it seems to value presaturating before expanding. But to always be well marcoed and to make the combat decisions it is making is very impressive. Obviously it’s through testing 200 years of possible ways to interpret these scenarios. But things like the ramp block are so fascinating because there’s so many ways to build a base and it’s not entirely obvious to a brute force situation why you would ever reach outside of building in base.
It also showcases how you have more options if you have better mechanics. It can win situations a worse mechanical player could not. Mechanics and strategy in humans are linked and in the game they are also linked. It’s not possible to separate them entirely because the level of depth to your game expands as your mechanics open up new avenues.
Didnt the bot mana played train for even more than 200 ingame years? I wonder if its capped in its potential. I mean, 200 years of playing and the bots havent figured out that you can distract someone by dropping immortals with a warp prism. Talk about a slow learner haha.
I badly want to see them let this bot grind much more and then let it play with half the apm.
Well for best results they should half the APM before they let it grind, no?