|
Mata posted:Until the deepmind team can accurately implement wrist pain, depression and disapproving parents, this alphastar farce proves nothing. What, no match fixing scandals?
|
#
?
Jan 25, 2019 12:42
|
|
|
|
| # ? May 9, 2024 17:27 |
|
|
They have a Terran DeepMind in development too, but it just spends all day on reddit bitching about protoss.
|
|
#
?
Jan 25, 2019 14:17
|
|
|
I for one don't want to see what an AI can do when bound by limitations like max APM. I thought the whole point of this was to make as good an AI as possible. If it's unbeatable and ludicrous with micro that's fine.
|
|
#
?
Jan 25, 2019 14:44
|
|
|
VelociBacon posted:I for one don't want to see what an AI can do when bound by limitations like max APM. I thought the whole point of this was to make as good an AI as possible. If it's unbeatable and ludicrous with micro that's fine. I was hyperbolizing a bit above but i would be disappointed if the AI had such good micro/macro that it didn't need to learn any strategy, tactics or decision-making at all, and just won with the same 1 base blink stalker build every game. That would seem to me like the deepmind team wasted their time on a game that wasn't very suited for what they were trying to do.
|
|
#
?
Jan 25, 2019 14:58
|
|
|
No Idra GG from the bot, not real enough.
|
|
#
?
Jan 25, 2019 16:06
|
|
|
Mata posted:I was hyperbolizing a bit above but i would be disappointed if the AI had such good micro/macro that it didn't need to learn any strategy, tactics or decision-making at all, and just won with the same 1 base blink stalker build every game. That would seem to me like the deepmind team wasted their time on a game that wasn't very suited for what they were trying to do. I guess I can see that perspective.
|
|
#
?
Jan 25, 2019 16:32
|
|
|
Imagine how powerful zerglings would be with perfect, optimal micro. Seems like the AI would easily destroy any zerg player in the early game every time, and it would never even need to learn how to go beyond that. Although I guess it would if it were playing against an equally matched AI opponent and maybe you'd get some really nutty ling/bane skirmishes
|
|
#
?
Jan 25, 2019 16:53
|
|
|
Life was a perfect zergling micro bot. You can’t prove he wasn’t.
|
|
#
?
Jan 25, 2019 17:04
|
|
|
quote:Re. 1: I think this is a great point and something that we would like to clarify. We consulted with TLO and Blizzard about APMs, and also added a hard limit to APMs. In particular, we set a maximum of 600 APMs over 5 second periods, 400 over 15 second periods, 320 over 30 second periods, and 300 over 60 second period. If the agent issues more actions in such periods, we drop / ignore the actions. These were values taken from human statistics. It is also important to note that Blizzard counts certain actions multiple times in their APM computation (the numbers above refer to “agent actions” from pysc2, see https://github.com/deepmind/pysc2/blob/master/docs/environment.md#apm-calculation). At the same time, our agents do use imitation learning, which means we often see very “spammy” behavior. That is, not all actions are effective actions as agents tend to spam “move” commands for instance to move units around. Someone already pointed this out in the reddit thread -- that AlphaStar effective APMs (or EPMs) were substantially lower. It is great to hear the community’s feedback as we have only consulted with a few people, and will take all the feedback into account. The problem is not the raw APM. The problem is that AlphaStar almost never mis-clicks or worries about mis-clicking.
|
|
#
?
Jan 25, 2019 19:15
|
|
|
VelociBacon posted:I thought the whole point of this was to make as good an AI as possible. But that's obviously not interesting, computers doing basic computer actions faster and with more precision than humans is expected and boring. And if the AI is mainly winning because it pairs 'okay' level strategy with absurdly strong mechanics, that's...well it's an improvement on the generally terrible strategy and decisionmaking of past bots, but it doesn't mean your computer is pro-level at strategy and tactics, which is the thing they're trying to make AI good at. The whole point of DeepMind is figuring out the strategy and tactics type of stuff, not being able to click super fast. The ideal here is micro that's at roughly the same level of the humans they're facing, that way if they win you know it's because of superior tactics and strategy, rather than brute force. Sure, we admire humans with really good micro, and we do so the same way we admire humans who run super fast, but that doesn't mean we're impressed by a car going five times faster.
|
|
#
?
Jan 25, 2019 21:28
|
|
|
I'd still find it interesting to see what the absolute highest level Starcraft that could be squeezed out of a computer would look like. It may not be as interesting scientifically, but it'd be interesting from the perspective of the game Starcraft 2. I guess you could keep the showmatch version screen limitation, but I feel if you remove the apm cap that the effect of that would diminish anyway. But hey, I'd also be super interesting in watching a big 4v4 Archon mode game with the best players there are. In what we saw, it's not perfectly top human level either. It's obvious that AlphaStar has micro skills far superior to any human player. Not only can it rely on this superior micro, it can go into battle knowing it will micro this perfectly, it doesn't need to take into account the risk of making a mistake in it. It loves Stalkers yet doesn't really like Immortals, and I think that's probably because at that level of micro the Stalkers just win. They survive so much, where an Immortal could have been chased and focused down. the Stalker's speed and blink make it stronger if you have better micro, where the Immortal's slower speed and shield makes it weaker if your opponent has better micro. Remember that this AI has been trained to play against itself, so it's expecting the same micro from it's opponent. The AI kind of making a mistake (and MaNa capitalizing on that very well) is interesting too. There were some other things dismissed by the casters as bad play (going up the ramp in the first game against TLO, overprobing) where I'm not so sure. Even though I agree with Rotterdam that it's possible to make a bad move and still win because your opponent messes up his response, it's hard to say. You can't say something is a good decision just because it won, but I don't think you can call it bad without seeing it lose either. The AI made the decision to come up on that ramp knowing it could back it up with unbelievable micro. Even if it'd be forced to retreat, it would be able to do so exceptionally well. And I feel the overprobing probably has some uses as well in the AI vs AI meta. Maybe it's to preemptively replace harass losses, or as a safer way to build up an expansion. The AI is not infallible and if you give a pro level player a full day with it they could probably poke holes in it and start beating it most/all of the time, but I'll bet there's also some things that the pro scene could learn from this AI.
|
|
#
?
Jan 25, 2019 23:19
|
|
|
I just want to see Korean pros get dunked on by a perfected ling rush. Over and over.
|
|
#
?
Jan 26, 2019 00:21
|
|
|
MisterZimbu posted:They have a Terran DeepMind in development too, but it just spends all day on reddit bitching about protoss. The in development zerg deepmind meanwhile will only ling flood regardless of what they try. Occasionally it will banebust if the flood doesn't cut it, but it's still never produced a roach in 1200 iterative years
|
|
#
?
Jan 26, 2019 02:20
|
|
|
It knows they're supposed to be 1 supply.
|
|
#
?
Jan 26, 2019 02:26
|
|
|
I guess it knows more cheese than just 4Gate. https://www.youtube.com/watch?v=H3MCb4W7-kM&t=4603s
|
|
#
?
Jan 26, 2019 03:33
|
|
|
GSL Code S starts next Friday (EST)! And the old logo is back! Doesn't look like groups have been drawn yet but Scarlett is the only foreigner to qualify. https://www.youtube.com/watch?v=WD1pCC5_aRE 
|
|
#
?
Jan 26, 2019 16:03
|
|
|
Holy poo poo MC qualified. Did the big name foreigners try to qualify? Like serral?
|
|
#
?
Jan 26, 2019 16:13
|
|
|
VelociBacon posted:Holy poo poo MC qualified. Did the big name foreigners try to qualify? Like serral? I think just Scarlett and SortOf, plus a Chinese player called Cloudy. It sounds like there are some new region locking rules no one is supposed to talk about yet, from what I gather if you participate in WCS Korea you can't participate in WCS Challenger and vice versa? Would be good of Blizzard to actually announce the calendar and rules for 2019 I guess.
|
|
#
?
Jan 26, 2019 16:47
|
|
|
Are they going to continue on with deep mind so it can play all the matchups on different maps or are they done?
|
|
#
?
Jan 27, 2019 01:11
|
|
|
Ratios and Tendency posted:Are they going to continue on with deep mind so it can play all the matchups on different maps or are they done? From what I can gather they will expand the project to other races etc. I wonder if they took the approach of handicapping the micro to be somewhat closer to what humans can do, introducing small chance of error, misclick etc. whether the game would evolve strategies more useful and applicable to humans. Because it makes sense that it would hammer you with stalkers because a robostalker is a much better/different unit than the ones we use.
|
|
#
?
Jan 27, 2019 01:29
|
|
|
Hell yeah gsl
|
|
#
?
Jan 27, 2019 03:59
|
|
|
Byun is currently doing his last stream before beginning his two-year military service. Presumably he has to stop soon, as he's been at it for 18 hours, but who knows? https://www.twitch.tv/byunprime Yoshi Wins fucked around with this message at 04:23 on Jan 27, 2019 |
|
#
?
Jan 27, 2019 04:07
|
|
|
Yoshi Wins posted:Byun is currently doing his last stream before beginning his two-year military service. Dude went for at least 24 hours. I dropped out then. I think he may have added another hour even?
|
|
#
?
Jan 27, 2019 16:35
|
|
|
Playing the long game of "I can't serve or perform military duties with this crippling exhaustion and carpal tunnel, oh well!"
|
|
#
?
Jan 27, 2019 16:56
|
|
|
Man what a weird feeling that must be. I'll miss him for sure.
|
|
#
?
Jan 27, 2019 17:29
|
|
|
https://medium.com/@aleksipietikine...he-1702fb8344d6 Pretty good article on why the Deepmind exhibition was kind of BS. Fairly detailed and well-argued. Seems like all the commenters agree as well. I hope that the exhibition generally gets bad reviews, because I am interested in seeing them try to make an AI that doesn't win mostly through raw micro superiority. Go and chess players were impressed with Deepmind's go and chess AIs. I think Deepmind can do something interesting with Starcraft. They just didn't do it last week.
|
|
#
?
Jan 28, 2019 21:21
|
|
|
I'm pretty sure that the person who wrote that doesn't play any meaningful amount of starcraft and lol they used like 5x more space than they needed to make their point and all their images are literally the images deepmind released to show the relative limitations they imposed. Their central idea that spam clicking is the cause of high apm bursts is weird because the super high apm bursts are, almost without exception, production hotkey related as even people who can click very quickly are usually only in the 8-11 clicks per second range. The accusations that deepmind is being wildly deceptive and dishonest is just strange as they conducted this in about the most transparent way possible (down to even releasing a bunch of extra replays). Anyways the rest of the premise of the article is weird too and, imo, misses the point: they weren't making a bot that has the precise limitations of humans (otherwise it would just be a serral bot), but rather one that is in a similar general ballpark to humans. If you watch other AIs play, they are regularly at 5000-20000 sustained apm throughout games with nearly perfect micro. Alphastar did some great micro, but was by no means even close to perfect. Tbh most of the fights were won because the micro was good enough but the macro was so perfect that a warp in was never missed. In any event, by capping the bots apm at a not insane (but still high) amount it emphasizes the decision making and overall strategy as opposed to individual fight tactics. That the bot was almost always performing actions in just one part of the map at a time (with near instant macro cycles interspersed) was probably the biggest limitation. The 350ms delay in reaction-processing-time was substantial as well, that's 150ms slower than most pro players (and if we see a zvz version of the bot, I'm sure that will be relevant early game wrt baneling micro/fights). Besides, they were straight forward about not presenting this as some definitive proof of really anything and tbh I think the results were surprising and interesting enough even in the peculiar context and I'm just really curious to see where it goes. It's remarkable that the actual agents were only a week or two in development E: imo the most significantly unbalanced thing isn't even the effective micro or epm bursts or anything else, it's the very near zero percent error rate in actions. It's actually this that made the games reminiscent of Serral's playstyle in that he is so consistently good at turning an engagement to his favor and then cementing the advantage he gains and ultimately snowballing it into a win. Like that ability is by definition human, just not really relevant to 99.99% of starcraft playing humans. E2: of note, and the author misses this, the primary spike in ms of reaction time is at 190, which is a fast human reaction time. The deepmind team did quite a bit to get the playing field at least marginally level. Herstory Begins Now fucked around with this message at 22:51 on Jan 28, 2019 |
|
#
?
Jan 28, 2019 22:27
|
|
|
I really hope they add this AI to the in game AI list as an option or something because it would be really fun to play against.
|
|
#
?
Jan 28, 2019 23:06
|
|
|
Herstory, you bring up some good points about the delay. That's definitely an important limitation that helps level the playing field between man and machine. But I also still think that AlphaStar played in a way that is physically impossible for a human, and I agree with the author of the article that Deepmind's reps repeatedly implied that a human could do everything AlphaStar was doing. I think we're looking at this from very different perspectives when you dismiss a bot with fully human limitations as "just a Serral bot." Actually, I would be much more impressed if they made a bot that I couldn't distinguish from Serral. I think it would be a lot harder to make such a bot. It would have to be a lot better than I believe AlphaStar is at evaluating the game state. It's interesting to hear that other, weaker bots sometimes have up to 20,000 APM. That proves that Deepmind did accomplish something new. I still hope they jigger with the restrictions some more.
|
|
#
?
Jan 28, 2019 23:39
|
|
|
Yoshi Wins posted:Herstory, you bring up some good points about the delay. That's definitely an important limitation that helps level the playing field between man and machine. But I also still think that AlphaStar played in a way that is physically impossible for a human, and I agree with the author of the article that Deepmind's reps repeatedly implied that a human could do everything AlphaStar was doing. You should look into the BW bot tournaments. They're interesting in their own right and I'm pretty sure that the DM project took a lot of inspiration from that and Alphastar is 500x more human-like than anything from those bots. IMO 99% of what deepmind did was entirely possible for a human to do, albeit some of it only for 2-3 second bursts at a time maybe. Like the stalker micro, while incredible, wasn't literally perfect, but by managing to blink out 8/10 damaged stalkers instead of a human player blinking out 6/10, the more effective trade quickly changes the efficiency of the trade. That said, there's also a level of learning that pros would have to do against machines in order to have the same ability to visually pre-judge if a fight is worth it vs AI-microed units or human microed units. Mana generally has been a player who will take extremely close engagements, whereas someone who only engaged when in a more clearly favorable situation likely would have done better. Herstory Begins Now fucked around with this message at 01:24 on Jan 29, 2019 |
|
#
?
Jan 29, 2019 01:18
|
|
|
Yeah I would love to see unrestricted sc2 bots vs each other.
|
|
#
?
Jan 29, 2019 01:21
|
|
|
Yoshi Wins, I disagree in that I think the demonstration was quite impressive. So the issue is that if the AI has abilities in terms of faster more precise micro, and seeing the entire (visible) map at once, its most effective strategies are different than those a human would evolve. Since it has, in effect, super stalkers, super disruptors and comprehensive vision, it will use those units differently and value them differently than we would, and multitask differently. What impressed me, if I understand the presentation properly, is that the AI's play algorithm evolved from an evolutionary soup of combat with little more to instruct its principles than the rules and conditions of the game, and learning from games. If you wanted to see it evolve strategies that would be useful to human players it would be interesting if you altered its stats to be more similar to a human in terms of capped out max burst APM, one screen camera (as they did in the final demo game), introduced a slight inaccuracy into click targeting and calculations of range etc. similar to what a top human player might have, and THEN started the combat/evolution process, you might see some very interesting strats that could be properly utilized bya human player. It would be a different project though. I don't think it's fair to dismiss the demo as worthless just because Mana was able to abuse the last version of the agent by yo-yoing its F2 army around with a warp prism. With more trials and evolution time I'd think it SHOULD learn to respond properly to something like that, and if not, points out a limitation in the AI that needs to be addressed.
|
|
#
?
Jan 29, 2019 01:45
|
|
|
Zwabu posted:...if you altered its stats to be more similar to a human... ...you might see some very interesting strats that could be properly utilized bya human player. It would be a different project though. Yes, this is exactly what I want to see! This would rule. Zwabu posted:I don't think it's fair to dismiss the demo as worthless just because Mana was able to abuse the last version of the agent by yo-yoing its F2 army around with a warp prism. With more trials and evolution time I'd think it SHOULD learn to respond properly to something like that, and if not, points out a limitation in the AI that needs to be addressed. Worthless would be going a bit further than I'm going. I just didn't find it all that interesting. But it did do some things that seemed pretty strategically cool for a bot. I remember a game where Mana took his army out, lost an engagement, and tried to escape with the remainder of his force, and the bot did a pretty darn good job hunting it down through the fog of war. That was my favorite part of the demonstration (I didn't see every game). Anyway, I'm glad many people here appreciated it and have been following Starcraft bot development more than I have. I'm sure we'll see more of AlphaStar in the future, and I 100% expect it to be better at everything as time goes on.
|
|
#
?
Jan 29, 2019 04:31
|
|
|
The fifth game where it beat Mana it proxied a robo and shield batteries and stargate near Mana's natural and did an prism/immo/phoenix/gateway proxy rush that was impressive as hell.
|
|
#
?
Jan 29, 2019 04:38
|
|
|
OK, I missed that one and it sounds dope as hell. I'm gonna look it up.
|
|
#
?
Jan 29, 2019 05:46
|
|
|
Yoshi Wins posted:OK, I missed that one and it sounds dope as hell. I'm gonna look it up. https://www.youtube.com/watch?v=H3MCb4W7-kM&t=4603s
|
|
#
?
Jan 29, 2019 05:48
|
|
|
good post on reddit from one of the guys who runs the AI tournamentsquote:Hi guys, https://www.reddit.com/r/starcraft/comments/akk3mx/about_alphastar/ E: also 
Herstory Begins Now fucked around with this message at 09:59 on Jan 29, 2019 |
|
#
?
Jan 29, 2019 08:38
|
|
|
OK, I retract my opinion that the AI was uninteresting. I still agree with the charges that they were a bit deceptive about how they presented the machine’s limitations, particularly with respect to APM statistics, but I think they just needed a different script during the broadcast. Also, that fifth game was excellent. Thanks, Zwabu.
|
|
#
?
Jan 29, 2019 12:33
|
|
|
All effort posting aside, dude, the ai beat a fuckton of immortals with stalkers, that is completely insane even if nothing else happened that day E: the video of Mana talking about the whole experience is good, too https://www.youtube.com/watch?v=zgIFoepzhIo Herstory Begins Now fucked around with this message at 13:18 on Jan 29, 2019 |
|
#
?
Jan 29, 2019 13:13
|
|
|
|
| # ? May 9, 2024 17:27 |
|
|
Herstory Begins Now posted:The 350ms delay in reaction-processing-time was substantial as well, that's 150ms slower than most pro players (and if we see a zvz version of the bot, I'm sure that will be relevant early game wrt baneling micro/fights). Human reaction time starts getting much slower when you start adding in decision making, the necessity of different reactions depending on the stimulus, needing to click on a different area depending on the position of the stimulus, the starting frames of an animation not being immediately visible or obvious to a human like it is to a computer reading memory directly, whether or not the human is focusing on the stimulus or something else, a bunch of other noise going on at the same time, and so on.
|
|
#
?
Jan 30, 2019 02:24
|
|
