EPTVideo should be free,
https://blizzcon.com/en-us/watch?v=5db3b1d7b3701b837dee0e8e&eventYear=2019
AlphaStar reaches GM in all three races - Page 6
| Forum Index > SC2 General |
|
franzji
United States581 Posts
Video should be free, https://blizzcon.com/en-us/watch?v=5db3b1d7b3701b837dee0e8e&eventYear=2019 | ||
|
CyanApple
48 Posts
On November 03 2019 05:15 youngjiddle wrote: Lots of details and videos on the pylon show from blizzcon. Video should be free, https://blizzcon.com/en-us/watch?v=5db3b1d7b3701b837dee0e8e&eventYear=2019 DeepMind & TLO enter at around 22 mins. For me this was a very insightful interview & discussion of AlphaStar. For the first time I got the feeling, they have the right approach to training AlphaStar and that it is able to learn in an almost human way. Realizing that, AlphaStar becomes so much more valuable as we can learn about ourselves by observing Alphastar. Due to that, it becomes even more important to keep the project running and not stop at this point, where it becomes interesting. The most important insight here was: AlphaStar becomes smarter/more strategic, when its mechanical abilities are weaker. I'm very glad TLO could adress this during his work for DeepMind. This is something, I feel like, bothered the community the most in previous testruns. I would go even further to limit e.g. its APM (lower than pros), accuracy and, if feasible, add variance to all parameters affecting its accuracy. This could lead to an overall more robust/advanced gameplan, as it can't rely too much on some microcalculation and -corrections of its gameplay. Exactly knowing when to take an engagement and when not to, could limit the exploration of possible actions AlphaStar takes, thus missing out on learning something new, something that gives it an edge, eventhough it e.g. takes a devastating Disruptor hit (or maybe teaches it to prevent the enemy from building them in the first place, by scouting). Another interesting thought popped up, when I heard TLO explain, that AlphaStar became more patient. Generally this is expected from Deep Learning AI after following its success in chess and Go. The AI seemed to try to keep its winning percentage as close to 50% as possible, but be very certain that that is enough. It sacrificed at certain points, just to get ahead at the other side of the board. For the first time though, in StarCraft, this behaviour was simultaneously described as human-like. I think with the limited information players' have in StarCraft, they feel more comfortable building their advantage further (usually in economy or tech), rather than pressing the issue. It gives me the impression, that the highest level of StarCraft today is already quite close to the optimal way to play it. While in chess and Go you get the feeling high level AI could at some point just throw out an inhumane move and turn the game on its head. AlphaStar reaching GM in this game, that favours human intelligence much more, is for me then an even more incredible feat. I want to see AlphaStar improving to test the bounds of StarCraft. On a side note, adding exploiter agents is probably a very elegant/efficient way to introduce disturbances to AlphaStars training period, given you have the league system already. But to me this gives off the feeling of a very supervised learning, as you would probably have to somehow specify the cheese tactics, exploiter agents use. Thinking about this and AlphaStars general inability to adept to previously unknown situations made me wonder about a different way of training the AI, in a more human way. When we first start StarCraft, we don't jump in a regular game with hopes of winning it, neither do we learn unit interactions and purposes just by imitating other players' games. We focus on a subpart of the game and try to improve in certain areas, thinking of certain scenarios or achievements. First understanding different parts of the game separately to then combine them in a regular match. What better way to imitate this kind of approach than playing arcade training maps? You can learn to spread creep, you can learn to hit your inject cycles, place buildings to improve defense, you can play with limited available units, you can defend cheese scenarios. If AlphaStar could train on a gradually expanding actionspace and explore it freely (constraining the effective action space of a high level scenario, by previously training on a lower level scenario), would not just make it more independent of human supervision, but allow it to explore inhumane actions as well, while keeping the number of possible/promising actions small. This is exciting, after the first testruns I was quite disappointed in the approach and the results, but you really adressed the pressing matters in a very successful way. This topic deserves more attention and I can only call to the community to stay more involved and not get discouraged by minor setbacks. Give it relevance by giving it attention! | ||
|
necrosexy
451 Posts
On November 03 2019 05:15 youngjiddle wrote: Lots of details and videos on the pylon show from blizzcon. Video should be free, https://blizzcon.com/en-us/watch?v=5db3b1d7b3701b837dee0e8e&eventYear=2019 TLO making vipers vs phoenix lol no nydus counterplay no infestors i don't understand why Artosis is being so disingenuous. he knows that comp is dogshit against any decent zerg lategame | ||
|
sneakyfox
8216 Posts
| ||
|
necrosexy
451 Posts
On November 04 2019 04:48 sneakyfox wrote: Alphastar beat Serral 4-1 in ZvZ: https://twitter.com/LiquidTLO/status/1190779241564000256 https://twitter.com/LiquidTLO/status/1190796307700387841 Serral couldn't use his own gear nor set up his hotkeys | ||
|
sneakyfox
8216 Posts
On November 04 2019 05:42 necrosexy wrote: Serral couldn't use his own gear nor set up his hotkeys lol ok, did he also have to play with his feet?  | ||
|
necrosexy
451 Posts
On November 04 2019 06:15 sneakyfox wrote: lol ok, did he also have to play with his feet? maybe Alphastar did vs Florencio | ||
|
shabby
Norway6402 Posts
On November 02 2019 10:32 cha0 wrote: Making this general claim of cars being better drivers than humans is one of the most laughable things I've ever seen in any AI thread. You realize not all cars has to be better than all drivers? In general, accidents would go way down with automated cars in the current state. Overall, they are better drivers. The most common cause of (lethal) accidents, is microsleep, just eradicating this alone would make an enormous impact. Maybe you think that "Better" means faster on a race track, and not safer from A to B. | ||
|
deacon.frost
Czech Republic12128 Posts
On November 04 2019 15:14 shabby wrote: You realize not all cars has to be better than all drivers? In general, accidents would go way down with automated cars in the current state. Overall, they are better drivers. The most common cause of (lethal) accidents, is microsleep, just eradicating this alone would make an enormous impact. Maybe you think that "Better" means faster on a race track, and not safer from A to B. Automated cars have issues. For example a car with optics for over 100k USD isn't able to distinct a child from a dust bin (the small one) in winter. Because all the fancy clothing makes children look like small walking dust bins. But the AI says dust bins don't walk. AFAIK this still hasn't been resolved, but who would care, the chances are small and in the end it's the drivers responsibility to take over the car in case of SHTF. Also global warming will solve this once and for all (most of the time the kids don't run under a car in winter, it's highly specific, but still it's an issue) And that's just one issue, the driving software have many issues and not everything can be solved. And I am still ignoring the two biggest issues - legal one(who's responsible for accidents in a fully automated car which's supposed to be better than human - because when something works for most of the time people tend to start to ignore the thing thus you can't say - the driver is responsible, driver will be watchin a movie on his phone, this is just a scapegoat) and moral one(in some accidents the car may be able to decide who to harm - what should be the base for such a decision?) | ||
|
CyanApple
48 Posts
I think it would already help for the casters to specify exactly which replay they are casting (by demo name in the title, I guess?). Then I would love to see a collective overview of which replays were already cast, maybe with remarks on whether the game was rather special or not Maybe on the liquipedia page of AlphStar or something. I could also start a thread, in which youtuber/streamer could answer with a compact overview of their own casts? How do you feel about that? | ||
|
cha0
Canada501 Posts
On November 04 2019 15:14 shabby wrote: You realize not all cars has to be better than all drivers? In general, accidents would go way down with automated cars in the current state. Overall, they are better drivers. The most common cause of (lethal) accidents, is microsleep, just eradicating this alone would make an enormous impact. Maybe you think that "Better" means faster on a race track, and not safer from A to B. Would you stake your life on this automated car on a wet road during a snowstorm in the night? Cars may be better drivers in certain ideal conditions, but they are still far from prepared to drive in all situations. | ||
|
WombaT
Northern Ireland23755 Posts
On November 05 2019 17:02 cha0 wrote: Would you stake your life on this automated car on a wet road during a snowstorm in the night? Cars may be better drivers in certain ideal conditions, but they are still far from prepared to drive in all situations. Possibly, depends how they’ve got the tech down so far. Sure there’s a lot of reasonable skepticism against automated cars and plenty of issues to be corrected. We’re striving for close to perfection in terms of safety when the current state of affairs is far, far from that. I’d not stake my life on crossing the road with a driver who’s looking at a meme on their phone, or a drunk or stoned on either. | ||
|
fishjie
United States1519 Posts
god if they actually figure out how to generalize more and more, say goodbye to everybody's job. GG NO RE | ||
|
FFW_Rude
France10201 Posts
| ||
|
Ej_
47656 Posts
On November 06 2019 01:11 fishjie wrote: This is a terrible idea god if they actually figure out how to generalize more and more, say goodbye to everybody's job. GG NO RE Will the robots take progamers' jerbs?! | ||
|
BronzeKnee
United States5212 Posts
On November 04 2019 04:48 sneakyfox wrote: Alphastar beat Serral 4-1 in ZvZ: https://twitter.com/LiquidTLO/status/1190779241564000256 https://twitter.com/LiquidTLO/status/1190796307700387841 Serral is playing against or as Protoss there... I don't even know what to say. Anyone who takes the time to watch the replays knows Alphastar is low to mid GM at best. I guarantee I crush it with cheese, it makes terrible decisions, Platinum level decision making. It struggles versus multi-prong attacks, and it is easily baited into traps. And it doesn't adjust on the fly. You bait it into a trap once... it will walk right into it again... and again... and again. Watch the longest TvZ game in the replay pack. The Zerg player makes it look stupid. Watch the replays. | ||
|
brickrd
United States4894 Posts
| ||
|
necrosexy
451 Posts
this is a fair appraisal | ||
|
Parser
Italy86 Posts
On November 06 2019 01:23 FFW_Rude wrote: Was Alphastar a blizzcon only thing or will we get AlphaStar playable in the future for everyone ? This is what interests me the most. I wonder if someone as informations about that, my searches have given no information at all about this possibility. | ||
|
heqat
Switzerland96 Posts
On November 08 2019 11:16 brickrd wrote: the fact that people get so heated and invested in angrily debating what league or MMR alphastar is so they can look smart criticizing it instead of just saying "wow, how cool that someone created this and that it can play competitive maps against pros" is really sad, another instance of miserable internet people shitting on someone else's accomplishment I think the problem is that DeepMind tends to overhype their results. I watched the Pylon show live event with TLO and it was all about how great is AlphaStar at SC2, etc. There would be less criticizing if they would spend some time talking about the limitiation of their SC2 agents which many people noticed (see the beastyqt video for instance). Of course, in the context of AI research it is a great achievement (which is what matters at the end), but in the SC2 context they should communicate more on the fact that SC2 is not mastered at a pro level (which was the case for Go and Chess). | ||
| ||
