Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepMind StarCraft II Demonstration [video]

twitch.tv

Post reply on HN
Print view View original

51–60 of 124 posts

Re: DeepMind StarCraft II Demonstration [video]

#51

Post by plaidfuji »

This is mad impressive, for sure. With most AI problems, I can at least comprehend the approach, and the necessary combination of models. Not so much here.

My first question is what is the input to the AI? Is it the raw pixel array of the display? Or does it get API-level readouts of what’s happening? Because implementing the CV just to segment the display output in real time is crazy enough. I would assume the latter.

I think this basically proves that any problem that can be exhaustively simulated is solvable now. This may mark a tipping point, as every problem for which simulations exist (essentially infinite labels) is solved - then the balance will tip back toward making faster and more accurate sims (think multi-scale first principles physics stuff).

Re: DeepMind StarCraft II Demonstration [video]

#52

Post by dwaltrip »

maerF0x0 wrote: post #32

Some of the commentary shows that the UI is a barrier . The fact we can understand the strategies, but cannot physically make it happen shows that at least some of the advantage is just the precision of the inputs. Could be a new way to play SC2 like games where we can better communicate our intentions to the game. For example, make a type of move action where the stalkers automatically retreat and stop to fire, inst…

I was just daydreaming the other day about an RTS where the UI controls and inputs are fully customizable by the users, for this exact reason. There could be even a community repository where the best known custom UIs are available.

Re: DeepMind StarCraft II Demonstration [video]

#53

Post by devilmoon »

I'm not proficient in Starcraft, but if they have to make Pros play against 5 different agents for a Bo5 is it because the same agent would merely repeat the same game overall and the human player would be able to see through its strategy?

AFAIK at least for Go/Chess DeepMind wasn't handpicking agents to send against human opponents, but it was simply a trained agent who would try its own strategy and respond to the opponent, in this case isn't it like the single agent only ever plays a single type of meta strat? If so, I think this is a bit less impressive than what I had predicted.

Re: DeepMind StarCraft II Demonstration [video]

#54

Post by devilmoon »

kmnc wrote: post #45

Very impressive... but it seems like the AI relies entirely on abusing blink stalkers which with perfect micro is basically impossible to counter. It is no surprise it can crush pros when it has perfect timing and zero mistakes in using these units. I think the coolest thing is how the play of AI mirrors a similar style to how pros have developed (Micro harrasing early, early expansions, a very good understanding of…

>I think the coolest thing is how the play of AI mirrors a similar style to how pros have developed

I mean, the agents have learned the game from pro replays so to me it seems obvious it would evolve to play the same strats human pros use.

Re: DeepMind StarCraft II Demonstration [video]

#55

Post by gambler »

kmnc wrote: post #45

Very impressive... but it seems like the AI relies entirely on abusing blink stalkers which with perfect micro is basically impossible to counter. It is no surprise it can crush pros when it has perfect timing and zero mistakes in using these units. I think the coolest thing is how the play of AI mirrors a similar style to how pros have developed (Micro harrasing early, early expansions, a very good understanding of…

Stuff like that is why StarCraft is not a very good game to test AI on. It's good for publicity, because it is well-known and has a pro scene so you can claim to "beat humans", but it's too complex in some ways (ruleset) and too simplistic in others (mostly just killing stuf). APM and micromanagement are huge, while long-term strategies are fairly limited compared to many other strategy games.

With all of this, it's really hard to see how much/which parts of the game the AI really "understands".

I'd love to see an AI that can play MAX or MoO from raw pixel inputs, even if it sucks at it.

Re: DeepMind StarCraft II Demonstration [video]

#56

Post by degenerate »

maerF0x0 wrote: post #32

Some of the commentary shows that the UI is a barrier . The fact we can understand the strategies, but cannot physically make it happen shows that at least some of the advantage is just the precision of the inputs. Could be a new way to play SC2 like games where we can better communicate our intentions to the game. For example, make a type of move action where the stalkers automatically retreat and stop to fire, inst…

This thought process (of the UI limiting the player) is a valid point when we are talking about Human vs. AI strength and weaknesses in RTS games. However, when pitting one human player against another, the UI limitation actually adds strategic depth to the game. Unlike pure strategy games like chess, in Starcraft when you have to move, attack, move, attack, it puts a physical burden on the player. This physical burden to work around the limited UI allows one player to outperform the other on both a strategic and physical playing field, in real time. If both players were freed of their physical shackles, it would become a much more boring display of pure strategy, which eventually would be "solved" and the game would cease to be fun.

If you look at the difference between Starcraft 1 and Starcraft 2, the two most significant UI changes are the ability to select more than 12 units at a time, and the ability to "multi-cast" spells in sequence when more than 1 spellcaster is selected. These two small UI improvements greatly reduce the skill gap between an average player and a "pro" player in Starcraft 2, to such a degree that many Starcraft 2 pro players went back to Starcraft 1, where their fast reflexes give them a much greater advantage against other players due to the limited UI. So an improvement to the UI would help human players beat AI, but it takes a strategic element out of the game when considering human vs. human.

Re: DeepMind StarCraft II Demonstration [video]

#57

Post by gamegoblin »

Buttons840 wrote: post #47

Twitch.tv chat is full of shit-posting no doubt, but I did find some of the comments amusing: - "You only have 50 GPUs in 2019, LuL" - "AlphaStar is a cyberbully" - Many were claiming the AI had "1500 APM", not sure where that idea came from - and lots more

There were moments during clutch blink micro where the onscreen APM number for AlphaStar exceeded 1000.

Re: DeepMind StarCraft II Demonstration [video]

#58

Post by cjbprime »

Veelox wrote: post #50
iamjaredwalters wrote: post #43

Earlier quoted context omitted.

TLO and Mana are good players but Im not sure they fulfill your criteria of top humans at X

According to a site [1] Mana is 19th. While that is probably overly high, I would say he is easily in the top 200 Starcraft 2 players in the world. I would also say that to get from beating a top 200 human to beating all humans is much smaller than from scratch to beating a top 200 human. [1] https://www.gosugamers.net/starcraft2/rankings

Aligulac's had Mana around 50th in the world consistently for a recent 12 month period, so I think that's a more reasonable claim, and I'm happy saying that the 50th best player is a "top player in the world".

Re: DeepMind StarCraft II Demonstration [video]

#59

Post by Zaita »

gambler wrote: post #19
Sweetness wrote: post #7

Earlier quoted context omitted.

Because it's still insanely impressive. I would have been blown away if the AI was merely at the platinum level. This is an AI playing starcraft. Starcraft!

>This is an AI playing starcraft. Starcraft! Yeah, it's not like computers were ever able to play StarCraft before. Oh, wait...

The AI normally picks a set strategy and attempts that. On harder difficulty the AI gets to cheat (more income per mineral collected). The SC2 AI is extremely primitive and at best Bronze level. The AI never really "reacts" to your actions, it's basically always on a set path. DeepMind is orders of magnitude better than the Starcraft AI.

FWIW cheating AI in RTS games is extremely common. Often they don't get fog of war, or have higher income rates. RTS AI is very very hard, so this is very impressive.

Re: DeepMind StarCraft II Demonstration [video]

#60

Post by unsigner »

dwaltrip wrote: post #52
maerF0x0 wrote: post #32

Some of the commentary shows that the UI is a barrier . The fact we can understand the strategies, but cannot physically make it happen shows that at least some of the advantage is just the precision of the inputs. Could be a new way to play SC2 like games where we can better communicate our intentions to the game. For example, make a type of move action where the stalkers automatically retreat and stop to fire, inst…

I was just daydreaming the other day about an RTS where the UI controls and inputs are fully customizable by the users, for this exact reason. There could be even a community repository where the best known custom UIs are available.

The best custom UI would have one button, "Launch DeepMind AI and let it win the game".

The intermediate UIs would be along the continuum between full micro and this.

Maybe developing these AI/UIs would be fun (until Google/DeepMind crushes you), but the game itself?

Post reply on HN