|
comedyblissoption posted: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. Here is a very cool article on Perception Action Cycles (PACs) in StarCraft 2. PACs are essentially how long it takes for a player to see something, take action, then move to attend to something else. It's a super good article if you want to nerd out hard. Before I dropped out of my PhD program I was going to try to springboard a dissertation off of that article. But yeah, 350 ms reaction time to something in sc2 is absurdly fast. The average human eye blink is about 200 ms, for context. Synastren fucked around with this message at 05:22 on Jan 30, 2019 |
# ? Jan 30, 2019 05:18 |
|
|
# ? Jun 13, 2024 03:07 |
|
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. That Dijkstra quote feels relevant here: "asking if a computer can think is a bit like asking if a submarine can swim". Ultimately, do you care about the process or the outcome more? edit: it feels a bit like some people in the SC2 community are falling into a trap they always criticise in newbies. They're saying that mechanics somehow aren't "real" Starcraft, but "strategy" and "decision-making" are what actually count. No, sorry. Mechanics are a huge part of Starcraft. "If you've got more poo poo than him, just go kill him." It's nonsensical to say that an AI victory would be legit if the computer were better than a human at strategy, but that if it's better at mechanics instead it somehow doesn't count. Zephro fucked around with this message at 15:42 on Jan 30, 2019 |
# ? Jan 30, 2019 15:29 |
|
They didn’t play chess and go in ways that are physically impossible for humans. Humans could physically place pieces in the squares that the programs did. Edit: I want to reiterate that I never said the demo had no value and I retracted my view that it was uninteresting. My remaining complaint is that I agree with the criticisms that they implied it played in a way that was mechanically within reach for humans, which I do not agree with. I do understand and agree that the demonstration represented a huge leap forward in this field. Yoshi Wins fucked around with this message at 18:38 on Jan 30, 2019 |
# ? Jan 30, 2019 18:34 |
|
Yoshi Wins posted:They didn’t play chess and go in ways that are physically impossible for humans. Humans could physically place pieces in the squares that the programs did. Watching Mana's commentary is pretty interesting. He's very frank about the nerves, and the impact of having basically no meta-knowledge of his opponent. It makes me want to see him play more games against it, heh. Zephro fucked around with this message at 19:20 on Jan 30, 2019 |
# ? Jan 30, 2019 19:06 |
|
Sure, ultimately everything is physical, in a sense. I am not a dualist! What I care about is whether the AI developments are likely to have practical applications. I don’t care that a submarine doesn’t “swim”, because a submarine can do practical tasks that can’t be done otherwise. A Starcraft-playing entity has no practical purpose. It can’t explore the ocean floor, the way a submarine can. However, if the methods that developed the AI can be applied to developing other AIs that do solve problems, then those methods have a purpose. Chess AI development from the mid-50s until very recently contributed almost nothing of value to the field of AI in general. The computers just got better at a specific type of brute force calculation. Other problems did not get solved because computers got good at chess via brute force calculation. I believe, but cannot prove, that an AI that wins on a pretty even mechanical playing field is more likely to be the result of AI developments that can be applied to other problems than an AI that wins in large part because of mechanical advantages. I could be wrong about this, but I think it makes sense. One piece of evidence is that decades of chess AI work turned out to be useless. But many other pieces of evidence are missing, and my belief is more of an intuition than anything else. Edit in response to your edit: Absolutely. The physical computation processes between man and machine are different. What I want to see is AI that only had computational differences. I think it’s less convincing if it does things a human can’t do because our eye and finger movements can’t be that fast or precise. Does that make sense? Yoshi Wins fucked around with this message at 19:33 on Jan 30, 2019 |
# ? Jan 30, 2019 19:30 |
|
The AHGL (Day9's inter-corporate gaming league) is dead, but it already has a spiritual successor: the Corporate eSports Association! Organized mostly by people who had volunteered in the AHGL. Games this season are Rocket League, Starcraft 2, Dota 2, League of Legends, CS:GO, Rainbow 6 Siege, PUBG, and Magic the Gathering: Arena. I'd link to a website but they haven't revealed it yet! Team registration opens Feb 1!
|
# ? Jan 30, 2019 20:05 |
|
Zephro posted:I know what you're getting at, but actually they did play those games in ways that are physically impossible for humans. Deep Blue especially: it could mathematically assess millions of positions a second. Whatever a human is doing when they play chess, they're not doing millions of explicit calculations like that. A human brain is not physically capable of that in the way that a silicon chip is. It's a bit fuzzier with AlphaGo, but no human is implementing Monte Carlo tree search, for instance, and no human could do it at the speed AlphaGo does. The version that played Lee Sedol also used huge gobs of compute - I forget exactly how much but it was at least dozens of GPUs. Whatever a human is doing when playing Go is a very different physical process from this, if only because the brain consumes ~20 watts while a big pile of GPU consumes several kilowatts. A human couldn't do that without literally melting. Apparently AlphaGo used 1920 CPUs and 280 GPUs, which is pretty ridiculous in itself.
|
# ? Jan 30, 2019 23:24 |
|
There was an interesting video from a couple of SC AI designers who made the point that, 'look the question with making SC AIs isn't even can you make a bot that can beat a top pro, but can you even make a bot that can beat a person period. it wouldn't even make much of a difference if they just grabbed some diamond 3 player because even at that level they're going to figure out the bots weaknesses and adapt appropriately.' Beating a good human player period is an accomplishment, beating a top pro 5-0 is insane. Like there are plenty of bots that already have near perfect micro that couldn't consistently beat a human player (eventually diminishing to never beat as the weaknesses of the bot are figured out) Also it's ironic that they got Mana for the demo because his playstyle has always reminded me of what we saw out of Alphastar with the constant aggression and turning fights that looked dubious af into victories or favorable trades. Yoshi Wins posted:Sure, ultimately everything is physical, in a sense. I am not a dualist! I suspect you are wildly wrong about this and are speculating about stuff you don't know about because if you listen to the deepmind people talk, they explain that their entire purpose of the project in general and their choice of SC2 in specific is that it, by incorporating so much uncertainty, gives it vastly more real world AI development application than anything they've done previously. In any event, calling all that previous effort useless probably is a bit of a misnomer as well considering they developed AIs that can greatly exceed the highest human potential in complex activities. What they learned doing that then influenced how alphastar was built and you'll notice that the foundational elements of alphastar actually involved setting up fixed activities for the bot to do that did have binary-type optimal/vs non-optimal solutions (which ironically was how they taught the bot to micro effectively). Herstory Begins Now fucked around with this message at 04:45 on Jan 31, 2019 |
# ? Jan 31, 2019 04:34 |
|
I think we're not really all that far apart. I agree that Starcraft provides an opportunity to create an AI that deals with uncertainty and that this is valuable because it's more comparable to real-world situations than go or chess. I then said that I think if more mechanically limited AIs still beat top players, that would probably be even more likely to show the development methods would be valuable in real-world situations. That's because human mechanical limitations aren't relevant to some real-world situations. I did throw a lot of shade on chess AI development prior to AlphaZero. I said the brute force approach researchers (not Deepmind) took for a long time when developing chess programs contributed almost nothing of value outside of that very narrow topic. I thought that was true though. I read a scholarly article on this about 6 months ago. I think it was this one: https://www.ncbi.nlm.nih.gov/pubmed/22530382 It's paywalled now, and it's also from 2012. I realize reading one paper doesn't make me knowledgeable. Maybe the paper was wrong, is outdated, or I misinterpreted it. I think in your last paragraph you're saying there actually was a direct connection between brute force AIs, such as old chess programs, and the development of AlphaStar's micro abilities. Do I have that right?
|
# ? Jan 31, 2019 06:00 |
|
Yoshi Wins posted:I did throw a lot of shade on chess AI development prior to AlphaZero. I said the brute force approach researchers (not Deepmind) took for a long time when developing chess programs contributed almost nothing of value outside of that very narrow topic. I thought that was true though. I read a scholarly article on this about 6 months ago. I think it was this one: https://www.ncbi.nlm.nih.gov/pubmed/22530382 It's paywalled now, and it's also from 2012. I realize reading one paper doesn't make me knowledgeable. Maybe the paper was wrong, is outdated, or I misinterpreted it. I can't say I'm especially familiar with the development of old brute-force chess AIs, but wouldn't it have been valuable as progress in the field of data storage and manipulation? Tracking a couple billion potential game states and their relationships is no easy task, particularly when you need to extract your response in real time, and I wouldn't be surprised if the modern deep learning techniques have some roots in attempts by researchers to reduce the complexity of that data.
|
# ? Jan 31, 2019 08:05 |
|
isndl posted:I can't say I'm especially familiar with the development of old brute-force chess AIs, but wouldn't it have been valuable as progress in the field of data storage and manipulation? Tracking a couple billion potential game states and their relationships is no easy task, particularly when you need to extract your response in real time, and I wouldn't be surprised if the modern deep learning techniques have some roots in attempts by researchers to reduce the complexity of that data. The most common criticism of Deep Blue and similar engines is that because so much of the knowledge is hard-coded and domain specific it makes them very very fragile. It's not just that you can't teach Deep Blue to play checkers. It would struggle with even trivial variations like Chess 960, where the starting order of the non-pawn pieces is randomised. Something like Glinski's, which is played on a weird hexagonal board, would be impossible for it. Go is different because it's mathematically intractable. There are just too many possible moves for brute-force to be viable. So a lot of the excitement around AlphaGo was because DM had figured out a way around that using something that's sorta-kinda like the heuristic reasoning that human pros are assumed to use. Besides it just being a cool new avenue and a very public display of the power of those methods, there was also hope that the approach would be less brittle and more adaptable to other problems. That seems to be true as well: AlphaZero, the most recent incarnation, was able to teach itself superhuman competence at chess or Go or shogi without human supervision and purely through self-play. The question of how relevant this all is to the real world is definitely interesting. DM like to talk about how a program related to AlphaGo was able to cut energy use in Google's data-centres by 40%, and to boost battery life in Android phones. To me the most impressive real-world result so far is in protein folding, which is a computationally intractable problem that's hugely relevant to all kinds of biology. There was a computer contest for protein folding just recently at which DeepMind's program absolutely wiped the floor with everybody else and hugely advanced the state of the art: https://deepmind.com/blog/alphafold/ So if all DeepMind were doing was attacking increasingly hard games (and SC2 is harder than Go in a bunch of important ways) I'd be less impressed (though still pretty impressed). But it does seem like the talk of real-world applications has at least some backing to it. And they are definitely right in that SC2 is a new and relevant problem class because it involves hidden information and very delayed rewards, which are things AI has struggled with. How and if this translates into the real world I don't know, but with their record I'm definitely willing to give them the benefit of the doubt. Zephro fucked around with this message at 12:49 on Jan 31, 2019 |
# ? Jan 31, 2019 10:59 |
|
Great post! I didn’t know about their protein folding work. What a fantastic result.
|
# ? Jan 31, 2019 13:44 |
|
Don't think anyone posted groups but hey GSL starts tomorrow at 2300 EST / 1300 KST with Dark v Trust & Keen v Dear! Maybe it's just not having seen anyone play in months, but several of these I'm really not sure who gets out second.
|
# ? Jan 31, 2019 14:00 |
|
Yoshi Wins posted:Great post! I didn’t know about their protein folding work. What a fantastic result. I just watched Mana game 5 and what the hell, haha
|
# ? Jan 31, 2019 14:56 |
|
Antares posted:Don't think anyone posted groups but hey GSL starts tomorrow at 2300 EST / 1300 KST with Dark v Trust & Keen v Dear! Wait is that MC?
|
# ? Jan 31, 2019 17:43 |
|
Incy posted:Wait is that MC? The one and only.
|
# ? Jan 31, 2019 17:45 |
|
WCS Europe and America group stages are being cast today! Europe Americas https://www.twitch.tv/starcraft
|
# ? Jan 31, 2019 18:32 |
|
Synastren posted:WCS Europe and America group stages are being cast today! Thank you! e: Just me or is this stream quality super bad? Like the black point is way too high so you have complete black areas instead of dark grey and there seems to be a contrast filter applied over it which makes things worse? e2: Let's get our group picks in for RO32? Group A 1st: Dark 2nd: Dear Group B 1st: Classic 2nd: Trap Group C 1st: Stats 2nd: GuMiho Group D 1st: sOs 2nd:Solar Group E 1st: Maru 2nd: HerO Group F 1st: INoVation 2nd: soO Group G 1st: TY 2nd: Leenock Group H 1st: Alive 2nd: Rogue VelociBacon fucked around with this message at 19:08 on Jan 31, 2019 |
# ? Jan 31, 2019 18:58 |
|
Zwabu posted:The one and only. Mcanning would like a word
|
# ? Jan 31, 2019 19:02 |
|
If anyone is trying to figure out when the games are for their local time, this link should autodetect your time zone and tell you.
|
# ? Jan 31, 2019 19:42 |
|
loving the cynicism from Incontrol where he can't help state that while Mcanning tried to throw the game, toodming refused to let that happen and threw harder
|
# ? Jan 31, 2019 22:43 |
|
I've felt like that recently in some of my ladder games. Keep loving up but thankfully my opponent fucks up harder
|
# ? Jan 31, 2019 22:50 |
|
VelociBacon posted:e2: Let's get our group picks in for RO32? I prefer doing my predictions on a group-by-group basis. My predictions for GSL Group A: Dark > Trust Keen < Dear Trust < Keen Dark > Dear Dear > Keen Dark and Dear advance! It's worth noting that if Dark somehow faces off against Keen, I expect a very close set. We get Group B for WCS EU and AM, so I'll toss those predictions out, too, I guess! WCS EU Group B: Heromarine < Mana Shadwn > Skillous Heromarine > Skillous Mana > Shadwn Heromarine > Shadwn Mana and Heromarine advance. Mana will do something gross to Heromarine in game 1, which will tilt Gabe. Gabe will lose the series, then maybe even lose the first game against the lower bracket opponent. He'll remember how to play and crush the other two nerds after that, though. WCS AM Group B: Scarlett > PiliPili ExpecT < Epic PiliPili > Expect Scarlett > Epic Pilipili > Expect Scarlett and Pilipili advance. Scarlett will maybe drop a map.
|
# ? Feb 1, 2019 02:07 |
|
Are you sure you have confidence in Mana? The other players will know how to beat him. Just pick Protoss and play like AlphaStar.
|
# ? Feb 1, 2019 05:14 |
|
Yoshi Wins posted:Are you sure you have confidence in Mana? The other players will know how to beat him. Just pick Protoss and play like AlphaStar. Mana will tilt Heromarine into a win, and the other two players are complete unknowns. While Mana may not be the best pro, nor the most consistent, I'm fairly confident he will beat a couple of European ladder heroes. You're right in that it's a risky claim, though!
|
# ? Feb 1, 2019 07:55 |
|
VelociBacon posted:Group H I'd go along with most of these but do you rate Alive that highly right now? I haven't seen him play in a while but I would've favored Rogue/Zest. Otherwise agree except I have faith Scarlett can get out second and Impact has a decent shot.
|
# ? Feb 2, 2019 00:39 |
|
Antares posted:I'd go along with most of these but do you rate Alive that highly right now? I haven't seen him play in a while but I would've favored Rogue/Zest. I have 0 faith in Scarlett unfortunately, I'd really like her to do well. Zest I just don't think is that great relative to Kespa era.
|
# ? Feb 2, 2019 00:45 |
|
Zest is almost certainly better these days than Alive, who basically hasn't done anything of note in literally a year or two at this point.
|
# ? Feb 2, 2019 01:49 |
|
poo poo didn't even realise gsl started today
|
# ? Feb 3, 2019 00:39 |
|
Feeling pretty good about my predictions so far. I was wrong about Pili, but that's ok. And while I didn't get the opening matches for Group B EU right, I still called events (and who advances!) correctly! GSL Group B is anyone's guess except we all know Classic will get through. He might drop a map to Trap, because PvP is volatile. And while I think Trap is favored to get out in second, I expect TRUE to pull a dark horse and barely get through. Missed WCS EU Group C, so for Group D: Lambo skates through handily. I have no idea which of the three people in his group will make it through to get eliminated in the ro16, but I'm going to randomly pick Gerald because IDs which are actual names amuse me greatly. WCS America Group D: JimRising will make it out in first, Neeb in second. Neeb's PvP is his strongest matchup, and I'm sure JimRising has his fair share of ZvZ. I'm expecting Jim to win 2-1 in the winner's match, and for Neeb to smoke the other two guys pretty easily.
|
# ? Feb 3, 2019 04:31 |
|
Antares posted:Don't think anyone posted groups but hey GSL starts tomorrow at 2300 EST / 1300 KST with Dark v Trust & Keen v Dear! Starcraft was good today boys
|
# ? Feb 8, 2019 07:39 |
|
JIZZ DENOUEMENT posted:Starcraft was good today boys Zephro fucked around with this message at 14:20 on Feb 9, 2019 |
# ? Feb 8, 2019 16:13 |
|
Marinelord's mentality is completely shattered.
|
# ? Feb 8, 2019 18:16 |
|
https://www.twitch.tv/gsl in 3 minutes!! I'm suggesting that sOs will get out first, followed by Solar. ed: Game 3 set 1 was very good and I highly recommend it! Synastren fucked around with this message at 05:39 on Feb 9, 2019 |
# ? Feb 9, 2019 04:57 |
|
Super entertaining games tonight E: games delivered, holy poo poo Herstory Begins Now fucked around with this message at 13:26 on Feb 9, 2019 |
# ? Feb 9, 2019 05:51 |
|
Live now! i mean it will be serral and namshar, this group is like cheating Live later! Probably TLO and Seither? Silky can certainly make a play to get out too, but I know nothing about this Chinese terran. https://www.twitch.tv/starcraft
|
# ? Feb 9, 2019 16:03 |
|
Ugh that Solar vs Bunny game was the worst throw I've seen in years. You know the one.
|
# ? Feb 9, 2019 20:04 |
|
Zwabu posted:Ugh that Solar vs Bunny game was the worst throw I've seen in years. You know the one. It's always fun to see someone throw when you're not emotionally invested in them.
|
# ? Feb 10, 2019 04:09 |
|
Nervous posted:It's always fun to see someone throw when you're not emotionally invested in them. Zwabu and I are Z players so it hurt to watch.
|
# ? Feb 10, 2019 07:29 |
|
|
# ? Jun 13, 2024 03:07 |
|
Zwabu posted:Ugh that Solar vs Bunny game was the worst throw I've seen in years. You know the one. Bunny would have pulled it off anyway
|
# ? Feb 10, 2019 07:48 |