A series of exhibition matches between DeepMind's'AlphaStar' and StarCraft II pros TLO and MaNa ended in a one-sided victory for the artificial intelligence which went a combined 10-1 against its human opponents.
The matches (broadcast from in-game replay) were played under the parameters of Protoss vs. Protoss and on the map Catalyst. AlphaStar was also restricted in the speed of its actions to keep it in line with human pros, such as having a 350 millisecond delay time between perceiving information and issuing commands. The AI also played with an average APM (actions per minute) of 280, compared to 390 for MaNa and 678 for TLO.
Despite making a number of decisions that would be considered questionable by human pro standards—such as over-saturated minerals with Probes, making very large number of Disruptors (and inflicting self-damage with them), or aggressively attacking high-ground defenses—AlphaStar was able to overpower its human opponents. Some of its plays resembled that of extremely-high APM bots (inhuman speed), such as complicated flanking actions or extremely precise Blink micro.
TLO and MaNa faced a number of different AlphaStar 'agents'—versions of the AI that trained through matches against each other in an internal league (combined with initial, imitation learning from human replays). With accelerated training, the agents were able to accrue around 200 years of real-time StarCraft II training over 14 days. The agents in the demonstration showed different preferences for strategies—for example one opted for mass Disruptors while others chose mass Blink-Stalkers.
While AlphaStar defeated both TLO and MaNa 5-0 in previous recorded matches, MaNa was able take a victory in a final, live exhibition game against a new agent. 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. While this new agent still had an estimated MMR of over 7000, MaNa was able to defeat its mass-Stalker strategy with careful scouting and an overpowering army of Immortals.
While AlphaStar chose Protoss and map to expedite the training process, DeepMind said the process could be applied to any race. Deepmind previously achieved notoriety for developing the AlphaGo AI which defeated a number of top Go players in exhibition matches.
On January 25 2019 06:43 Poopi wrote: I think the title is kinda misleading. Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate
This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
I don't think we will ever get to a consensus that the AI outsmarted us at the game due to mechanical aspect of the game. A good number of those games were won due to completely inhuman stalker micro.
I'd like to know epms and not worthless apms for this, considering the AI was capable of microing stalkers in 3 different places at once and win vs mass immortals.
On January 25 2019 06:48 ArtyK wrote: I'd like to know epms and not worthless apms for this, considering the AI was capable of microing stalkers in 3 different places at once and win vs mass immortals.
Honestly I feel that StarCraft II is much more of a mechanics game than 'strategy' game than most people think, (it's more real-time connect-four than chess), so I also felt this demonstration wasn't quite as impressive as it could have been.
At least for my personal taste, the truly impressive way to hold his exhibition would be to have an AI that's actually considerably BAD mechanically but still crushes human opponents based on strategies and tactics (and, I guess, EXTREME efficiency of actions).
On January 25 2019 06:48 ArtyK wrote: I'd like to know epms and not worthless apms for this, considering the AI was capable of microing stalkers in 3 different places at once and win vs mass immortals.
On January 25 2019 06:43 Poopi wrote: I think the title is kinda misleading. Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate
This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
How is that insignificant? The raw interface didn't lose, yet the camera interface lost the first game its played. That means there is a huge actual difference between the two. Keep in mind that it's an estimated MMR, it's not like the AI played human ladder and only managed to go to 7300 MMR instead of 7500. It's the internal MMR from their internal league "The version of AlphaStar using the camera interface was almost as strong as the raw interface, exceeding 7000 MMR on our internal leaderboard."
On January 25 2019 06:43 Poopi wrote: I think the title is kinda misleading. Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate
This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
How is that insignificant? The raw interface didn't lose, yet the camera interface lost the first game its played. That means there is a huge actual difference between the two.
This is where I stop engaging you in discussion ^_^
On January 25 2019 06:43 Poopi wrote: I think the title is kinda misleading. Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate
This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
How is that insignificant? The raw interface didn't lose, yet the camera interface lost the first game its played. That means there is a huge actual difference between the two. Keep in mind that it's an estimated MMR, it's not like the AI played human ladder and only managed to go to 7300 MMR instead of 7500. It's the internal MMR from their internal league "The version of AlphaStar using the camera interface was almost as strong as the raw interface, exceeding 7000 MMR on our internal leaderboard."
Alphastar lost because Mana adapted and new what to expect going in. All you have to do is defend like a regular insane AI and prioritize ramp defenses. I bet Mana could do a clean sweep of 5-0 if he played again.
I would also tend to put more weight in the results (despite low sample size of the camera-restricted agent) than their estimated MMR, at least without information about how those estimations are made. They said on the stream they really didn't know how powerful the agents would be or even if they had selected the agents which would have the best chance against human players. They don't seem to have a lot of faith in these MMR estimations. It seems to just be the MMR from the internal leaderboard.
On January 25 2019 06:43 Poopi wrote: I think the title is kinda misleading. Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate
This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
How is that insignificant? The raw interface didn't lose, yet the camera interface lost the first game its played. That means there is a huge actual difference between the two. Keep in mind that it's an estimated MMR, it's not like the AI played human ladder and only managed to go to 7300 MMR instead of 7500. It's the internal MMR from their internal league "The version of AlphaStar using the camera interface was almost as strong as the raw interface, exceeding 7000 MMR on our internal leaderboard."
Alphastar lost because Mana adapted and new what to expect going in. All you have to do is defend like a regular insane AI and prioritize ramp defenses. I bet Mana could do a clean sweep of 5-0 if he played again.
Not really, I watched the replay and, even tho I don't know PvP much, I'm pretty sure AlphaStar had the game somewhat won after the successful harass (yeah, even after the adept counter harass from MaNa), it had far better income, and sufficiently higher army value to have time to build a good composition.
But it got stuck really bad on the warp prism harass, and when MaNa started to attack near the 2nd base it was really indecisive.
So it feels like the AI can have a good start because it's relatively easy, but isn't able to change plans according to current situation (2nd oracle was queued up during the harass, no split of stalkers in the main base when the warp prism got cornered).
Reading the article, I'm not sure if their approach can overcome these difficulties given more time.
This was interesting. Nevertheless right now AI is playing a different game. It has much higher effective APM, by far superior micro management and superior vision, it will not be distracted in terms of sight. That is why it won games. The strategies presented by AI were rather poor and reactions questionable. I would say that AI has still a long way to go before it beats any pro players on equal terms, as successfull AI must prove itself beat players mainly by adopting superior strategies.
On January 25 2019 06:48 ArtyK wrote: I'd like to know epms and not worthless apms for this, considering the AI was capable of microing stalkers in 3 different places at once and win vs mass immortals.
I'm pretty sure apm = epm for the AI
yeah that's what made sense to me but i'd still like to know for all 3 during the matches
I'm quite impressed by the decision making and overall gameplan the AI showed but after looking at the replays the micro its displaying is completely inhuman, its basically playing with 10-20% more supply in terms of how efficient its units are.
As many others are pointing out, there's not much to be impressed with while the AI is playing a different game. We already know that a microbot using an API can outmicro a human using a mouse. I look forward to seeing what they come up with in the future, but what I saw today was no more interesting than if they had been feeding the AI bonus minerals.
Also looking forward to detailed replay analysis, I expect that will reveal that "APM" is a very misleading metric to be using.
On January 25 2019 07:17 No_Roo wrote: As many others are pointing out, there's not much to be impressed with while the AI is playing a different game. We already know that a microbot using an API can outmicro a human using a mouse. I look forward to seeing what they come up with in the future, but what I saw today was no more interesting than if they had been feeding the AI bonus minerals.
Also looking forward to detailed replay analysis, I expect that will reveal that "APM" is a very misleading metric to be using.
I mean - what we saw today was FAR superior to any bot we've ever seen in terms of strategy/decisionmaking so I wouldn't call it not impressive. It's definitely not on the level that it has better decion-making than a pro though.
WCS
TLO and 
).
