EPTI also want to point out that Pro APM compared to AI APM is hardly comparable, since every action of a PC is deliberate, while alot of actions of pro players are also spam. Maybe thats just nitpicking though.
AlphaStar AI goes 10-1 against human pros in demonstration…
| Forum Index > SC2 General |
|
404AlphaSquad
839 Posts
I also want to point out that Pro APM compared to AI APM is hardly comparable, since every action of a PC is deliberate, while alot of actions of pro players are also spam. Maybe thats just nitpicking though. | ||
|
deacon.frost
Czech Republic12129 Posts
On January 25 2019 08:21 404AlphaSquad wrote: I was extremely skeptical in the beginning. However i was positively surprised by the result. It is clear that Deepmind has done an amazing job. I do want to point out though, that I am fairly certain that after multiple games, the Pros can figure out a way to defeat the AI easily by exploiting AI patterns. That said, I am eager to see how much Deepmind can improve on what they already have. I also want to point out that Pro APM compared to AI APM is hardly comparable, since every action of a PC is deliberate, while alot of actions of pro players are also spam. Maybe thats just nitpicking though. Pros should less spam and do more effective action. Shame on them! Learn from the AI while AI is learning from you, this is a two way street  | ||
|
No_Roo
United States905 Posts
Why converge on practically monobuilding stalkers? Well, this is a unit composition that becomes more attractive the more you can out control your opponent for one thing. It's a good strategy if you can click faster and more accurately than your opponent. | ||
|
alexanderzero
United States659 Posts
On January 25 2019 07:30 Rodya wrote: Exactly; from what I've glanced at the games this isn't very interesting. Perfect orb-walking and dropship/blink micro. If humans could do that then all dropships would be heavily nerfed and blink would be removed from the game. I'm sorry Google - no one is shocked nor cares about this. Sadly their marketing department will continue to push this bot down our throats while they tell themselves that they've revolutionized AI. There's plenty of room to debate the fairness of this test, and what the results mean, but you're off base when you claim that this isn't a shocking result and a breakthrough. A lot of people don't realize this, but writing an AI that is even capable of accurately evaluating the game state and producing a halfway competent strategy IS the notable result here. Previous RTS AIs were totally incapable of doing anything resembling a reasonable strategy, whether they had 200 APM or 10,000 APM. Think back to the match where the stalkers targeted down the immortal that had popped out of the robo. No programmer ever entered code that said immortals were strong and that they should be prioritized. The AI figured that much out for itself, along with everything else it did to get there. | ||
|
AKnopf
Germany259 Posts
 Of course there were factors benefiting AlphaStar (only one map, only one matchup, maybe higher eapm than humans can do), but this was just a presentation to show how neural networks can learn a very complex domain with reinforcement learning. It was not a tournament to decide the fate of humanity, so maybe we can just relax a little over such details. It was truly fascinating to see how the agents got better from the first match to the second one with another week (~ 200 years) of training. Artosis was spot on to highlight the great decision making (respect for ramps, disengaging). I look forward to what AlphaStar will be able to contribute to the game. Maybe pros will think about building more probes earlier. Maybe we could have a live evaluation during human tournament games to see who is ahead or what a good strategy would be from the current state. Such tools are used in chess to great effect and can really add to the casting of a tournament. Of course the most important thing about AIs (to me) would be the whole ethic complexity. We develop something that by many definitions is truly intelligent, yet we have no rules and no laws on how to interact with such machines. Who is accountable for mistakes of the AI would be one of those questions. | ||
|
Gfire
United States1699 Posts
On January 25 2019 08:35 No_Roo wrote: Why converge on practically monobuilding stalkers? Well, this is a unit composition that becomes more attractive the more you can out control your opponent for one thing. It's a good strategy if you can click faster and more accurately than your opponent. I'd argue not just when you have better control than your opponent but when you have better control in an absolute sense. Meaning even playing against other AI agents with the same mechanical abilities the Stalkers will be the best unit. Even counter units like Immortals, also controlled by the AI, will fail against Stalkers, right? With equal control on both sides, as long as that level of control is high enough. I'd be interested what units it would land on in other matchups. | ||
|
jalstar
United States8198 Posts
| ||
|
vesicular
United States1310 Posts
Like everyone, I'm excited to see how it fares in the future when not using a raw interface, while facing pros who are not off-race as well as how it does in non-mirror match ups. The future here is rather exciting. | ||
|
teamTV
France16 Posts
From what that I have seen from the stream and the replays : equivalent of 3/4 gold players playing archon vs GM :D. Not really fair. Give the AI a mouse and the same camera : different story. 10-1 is a misleading title indeed. | ||
|
Xamo
Spain885 Posts
Most agents build way more stalkers than any other unit (see statistics at deepmind's website), because it is the best one on their "hands". Attacking or defending with it is too good of a strategy. This limits the strategic depth of the agent, particularly it does not need to scout because nothing counters stalkers properly. To overcome this limitation, they could lower the average APM limit while training, or put an absolute limit to the instantaneous APM. Almonst forgot it: AlphaStar is a tremendous achievement even at this stage of its development. Best gaming AI ever, easily. I cannot wait to see what Deepmind brings next time. | ||
|
LHK
204 Posts
>While the first ten agents agents were effectively able to 'see' the entire map at once (NOT a maphack—more akin to a max zoom-out), the new agent was given restrictions to mimic human player's field-of-vision limitations during a game. So much of Starcraft 2 is screen placement and attention. Letting the Agent basically be able to interact with an entire map at one point really defeats the purpose in my opinion. It's like letting your opponent move their chess pieces whenever they want while you can still only move 1 per turn. Feels like in that regard the Agent wasn't even playing "Starcraft" at that point. That said, it's really cool and I'm happy for them.. It just feels like something is off with their approach. Showcasing an Agent with SUCH a huge advantage (arguably a core, fundamental aspect of SC2 being removed in its favor) vs a human player just doesn't feel like it means much to me. You can't handicap an agent enough to make up for this insane advantage. | ||
Waxangel
United States33571 Posts
On January 25 2019 09:02 LHK wrote: While this was cool, i'm really disappointed in >While the first ten agents agents were effectively able to 'see' the entire map at once (NOT a maphack—more akin to a max zoom-out), the new agent was given restrictions to mimic human player's field-of-vision limitations during a game. So much of Starcraft 2 is screen placement and attention. Letting the Agent basically be able to interact with an entire map at one point really defeats the purpose in my opinion. It's like letting your opponent move their chess pieces whenever they want while you can still only move 1 per turn. Feels like in that regard the Agent wasn't even playing "Starcraft" at that point. That said, it's really cool and I'm happy for them.. It just feels like something is off with their approach. Showcasing an Agent with SUCH a huge advantage (arguably a core, fundamental aspect of SC2 being removed in its favor) vs a human player just doesn't feel like it means much to me. You can't handicap an agent enough to make up for this insane advantage. I mean all of these handicaps to try and mimic human limitations will be flawed. You could even take this further to and say "well the AI should only be able to focus its 'eyes' on a X by Y size portion of a screen at a time, and require at least Z milliseconds to process that information." You'll never please everyone, and the best we're gonna get to is 'good enough.' | ||
|
xongnox
540 Posts
Camera decision (where to look) and minimap-guessing ( you see a dot on the minimap : what is it ? should you respond to it now or next ? ) is very important skills in decent human players. Furthermore, between game+screen+eyes+brain, their is a minimal delay to understand and react each time a player change the camera location. Then there is even delay with mouse control, etc. Still, very impressive results although expected coming from Deepmind. Next step should be more mechanical fairness (see the minimap and 1 screen, max 4 screens/second, etc, mouse cursor simulation, etc) before playing all races all MU, because it is already over-fitting the "full zoom-out" setting, as seen in the mass blink-surrounding play. | ||
|
jalstar
United States8198 Posts
| ||
|
swissman777
1106 Posts
Moreover, deepmind should really cap the peak apm, bc no player can go 1000 apm microing stalkers at three different fronts chipping away all the units little by little. That's just micro-bot doing its thing. I wish we could limit its apm so much so that AI is forced to beat the human players not by micro, but by sheer ingenuity and picking the right compositions. That will be the actual great milestone in AI imo. | ||
|
xongnox
540 Posts
On January 25 2019 09:11 Waxangel wrote: You'll never please everyone, and the best we're gonna get to is 'good enough.' Yep. But full zoom-out hack (while very convenient from a technical standpoint) is far away to "Good Enough" conditions but closer to obvious cheat. It is even close to a human-used cheat, because even human players play better with possibility to zom-out more. | ||
|
snakeeyez
United States1231 Posts
I still think what they have right here, and these results of a fair AI that is no cheating playing a full game at this level and winning is still a pretty amazing achievement. I would like to see more games where the pro players can look for more weaknesses, but the AI being completely unpredictable makes it harder to predict than a normal person Its too bad they cant transfer this AI to brood war so we could see a flash or jaedong play against it. I also think brood war is a more balanced time tested game than sc 2 | ||
|
swissman777
1106 Posts
| ||
|
xongnox
540 Posts
On January 25 2019 09:20 jalstar wrote: The last AI we saw was still very good and would have beaten other humans who aren't as good as Mana Yes but it is time to stop over-fitting the zoom-out. I guess as seen with the 11th game they understand that. So using the cheat-code version to get to 10-0 is a bit frustrating, ideally they should have waited for the more-legit version to be able to win games vs pro before the exhibition matches. | ||
|
swissman777
1106 Posts
| ||
| ||
