The Thread To Remind Me We're Living In The Future

Seen something interesting in the news or on the intertubes? Discuss it here.

Moderators: Zamfir, Hawknc, Moderators General, Prelates

User avatar
Liri
Healthy non-floating pooper reporting for doodie.
Posts: 948
Joined: Wed Oct 15, 2014 8:11 pm UTC
Contact:

Re: The Thread To Remind Me We're Living In The Future

Postby Liri » Thu May 25, 2017 7:19 pm UTC

LaserGuy wrote:AlphaGo, a Google-developed Go playing program, clobbered reigning world champion 2-0 in a best of three Go series.

I know for the first game he didn't "clobber" him, but it's still crazy cool.
He wondered could you eat the mushrooms, would you die, do you care.

User avatar
Zohar
COMMANDER PORN
Posts: 7547
Joined: Fri Apr 27, 2007 8:45 pm UTC
Location: Denver

Re: The Thread To Remind Me We're Living In The Future

Postby Zohar » Thu May 25, 2017 7:26 pm UTC


Great news for one particular kiwi. Maybe.
Mighty Jalapeno: "See, Zohar agrees, and he's nice to people."
SecondTalon: "Still better looking than Jesus."

Not how I say my name

User avatar
Liri
Healthy non-floating pooper reporting for doodie.
Posts: 948
Joined: Wed Oct 15, 2014 8:11 pm UTC
Contact:

Re: The Thread To Remind Me We're Living In The Future

Postby Liri » Thu May 25, 2017 9:33 pm UTC

AI creates - and names - paint colors

Read the intro two paragraphs and then scroll down to the bottom to see the colors.

Spoiler:
I haven't laughed so much in a while
He wondered could you eat the mushrooms, would you die, do you care.

User avatar
flicky1991
Like in Cinderella?
Posts: 630
Joined: Fri Feb 11, 2011 3:36 pm UTC
Location: London

Re: The Thread To Remind Me We're Living In The Future

Postby flicky1991 » Thu May 25, 2017 9:36 pm UTC

Liri wrote:AI creates - and names - paint colors

Read the intro two paragraphs and then scroll down to the bottom to see the colors.

Spoiler:
I haven't laughed so much in a while

Her whole tumblr has funny neural-network-generated stuff - recipes, video game titles, Pokémon...

EDIT: Had wrong Pokémon link, was missing the illustrations
any pronouns (mainly she/he)
----
Avatar from ArguablyDerek
----
Discord for Forum Games posters
(Please let me know if the link doesn't work)

User avatar
Liri
Healthy non-floating pooper reporting for doodie.
Posts: 948
Joined: Wed Oct 15, 2014 8:11 pm UTC
Contact:

Re: The Thread To Remind Me We're Living In The Future

Postby Liri » Thu May 25, 2017 11:19 pm UTC

flicky1991 wrote:
Liri wrote:AI creates - and names - paint colors

Read the intro two paragraphs and then scroll down to the bottom to see the colors.

Spoiler:
I haven't laughed so much in a while

Her whole tumblr has funny neural-network-generated stuff - recipes, video game titles, Pokémon...

EDIT: Had wrong Pokémon link, was missing the illustrations

Yeah, I found it after I read the article. Very cute. One of them linked back to that xkcd color survey that I totally forgot about.
He wondered could you eat the mushrooms, would you die, do you care.

mfb
Posts: 944
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby mfb » Fri May 26, 2017 2:49 am UTC

Getting to space is quite good for the first flight of a company.


I don't see how the network could generate reasonable names just based on existing names. It can infer associations like "green is this region in RGB space", but it has no way to come up with new names that make sense. If there is no "carrot" in the training set, the network has no idea this is a valid English word and it doesn't know the color of carrots. If there is "carrot" in the training set, the network cannot do anything with it (apart from inventing "carrot orange" - great).

User avatar
sardia
Posts: 5843
Joined: Sat Apr 03, 2010 3:39 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby sardia » Fri May 26, 2017 3:07 am UTC

mfb wrote:
Getting to space is quite good for the first flight of a company.


Pfft, anyone can go straight up. How much delta v did it have? How close did it get to going 7.8 km/s?

morriswalters
Posts: 6936
Joined: Thu Jun 03, 2010 12:21 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby morriswalters » Fri May 26, 2017 3:48 pm UTC

Don't diminish what was accomplished. Fewer than ten private concerns have actually launched anything.

User avatar
pogrmman
Posts: 337
Joined: Wed Jun 29, 2016 10:53 pm UTC
Location: Probably outside

Re: The Thread To Remind Me We're Living In The Future

Postby pogrmman » Fri May 26, 2017 5:09 pm UTC

This rocket is designed as an orbital rocket, so it should be able to reach orbit.

My guess is they had issues with their trajectory that wasted delta-v.

Even going suborbital is pretty impressive!

User avatar
sardia
Posts: 5843
Joined: Sat Apr 03, 2010 3:39 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby sardia » Fri May 26, 2017 5:33 pm UTC

pogrmman wrote:This rocket is designed as an orbital rocket, so it should be able to reach orbit.

My guess is they had issues with their trajectory that wasted delta-v.

Even going suborbital is pretty impressive!

Oh good, that's much better than I expected.

mfb
Posts: 944
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby mfb » Sun May 28, 2017 3:02 pm UTC

sardia wrote:
mfb wrote:
Getting to space is quite good for the first flight of a company.


Pfft, anyone can go straight up. How much delta v did it have? How close did it get to going 7.8 km/s?
We only have the information from the company, but apparently the first stage worked as designed, the second stage ignited, and the payload fairing was discarded properly, but at the end it didn't burn long enough to reach orbit ("did not quite reach orbit"), they are investigating why.

Apogee seems to have been 250 km, for a trajectory to orbit that probably means they were nearly there.

mfb
Posts: 944
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby mfb » Mon May 29, 2017 1:20 pm UTC

LaserGuy wrote:AlphaGo, a Google-developed Go playing program, clobbered reigning world champion 2-0 in a best of three Go series.
Now 3:0, and AlphaGo also won 1:0 against a team of 5 human players playing together.
Humans are no challenge any more.

wumpus
Posts: 501
Joined: Thu Feb 21, 2008 12:16 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby wumpus » Mon May 29, 2017 2:10 pm UTC

pogrmman wrote:This rocket is designed as an orbital rocket, so it should be able to reach orbit.

My guess is they had issues with their trajectory that wasted delta-v.

Even going suborbital is pretty impressive!


It took spacex four tries to get Falcon[1] into orbit. From the sound of it, they should have enough data to fix their problems even faster.

While electric engines seem rather bizarre (they are competing with turbines!), I'd imagine a reliable start/stop/restart sequence would be invaluable. I can understand the Falcon Heavy "asparagus" staging being abandoned as a plumbing nightmare, but it sound easier if the engines could simultaneously stop and start.

Liri wrote:
LaserGuy wrote:AlphaGo, a Google-developed Go playing program, clobbered reigning world champion 2-0 in a best of three Go series.

I know for the first game he didn't "clobber" him, but it's still crazy cool.


It isn't programmed to "clobber", just beat the opponent. Looks like the old chart needs updating: https://www.xkcd.com/1002/

mfb
Posts: 944
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby mfb » Tue May 30, 2017 4:38 pm UTC

The AI prefers a high chance to win with a small margin over more risky moves that could lead to a larger margin. If the AI wins with half a point, that does not mean the game was close.
wumpus wrote:Looks like the old chart needs updating: https://www.xkcd.com/1002/
We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I made a list here. The comic was published 2012.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Tue May 30, 2017 5:01 pm UTC

mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.

User avatar
sardia
Posts: 5843
Joined: Sat Apr 03, 2010 3:39 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby sardia » Tue May 30, 2017 6:11 pm UTC

Mutex wrote:
mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.

He appears to simplify strategy into a simple rock paper scissors matchup via rushing. Or it's an innovative new strategy that only computers can play. Probably a little of both.

commodorejohn
Posts: 961
Joined: Thu Dec 10, 2009 6:21 pm UTC
Location: Placerville, CA
Contact:

Re: The Thread To Remind Me We're Living In The Future

Postby commodorejohn » Tue May 30, 2017 6:48 pm UTC

mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

As long as Calvinball is on the list, I'm not worried.
"'Legacy code' often differs from its suggested alternative by actually working and scaling."
- Bjarne Stroustrup
www.commodorejohn.com - in case you were wondering, which you probably weren't.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Tue May 30, 2017 7:01 pm UTC

sardia wrote:
Mutex wrote:
mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.

He appears to simplify strategy into a simple rock paper scissors matchup via rushing. Or it's an innovative new strategy that only computers can play. Probably a little of both.

Who is "he"?

User avatar
sardia
Posts: 5843
Joined: Sat Apr 03, 2010 3:39 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby sardia » Tue May 30, 2017 7:15 pm UTC


User avatar
heuristically_alone
Posts: 169
Joined: Sat Apr 09, 2016 7:43 pm UTC
Location: 37.2368078 and -115.80341870000001

Re: The Thread To Remind Me We're Living In The Future

Postby heuristically_alone » Tue May 30, 2017 8:02 pm UTC

mfb wrote:
LaserGuy wrote:AlphaGo, a Google-developed Go playing program, clobbered reigning world champion 2-0 in a best of three Go series.
Now 3:0, and AlphaGo also won 1:0 against a team of 5 human players playing together.
Humans are no challenge any more.


Best line from the article:
"Last year, I think the way AlphaGo played was pretty close to human beings, but today I think he plays like the God of Go," Ke said after the game.
Bow gifted by adnapemit.

Beatrice wrote:The world is quiet here.

Tillian wrote:sig'd

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Tue May 30, 2017 10:52 pm UTC

Mutex wrote:
mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.


I find it hard to believe that StarCraft is a frontier. If an AI is allowed to execute at maximum APM, then it would utterly destroy humans in basically every setting.

A huge part of StarCraft is about Micro, and computers will always execute Micro better than a Human. The main reason why AIs aren't programmed to take advantage of huge APM-tricks is because they're not fun to play against. For example, "Perfect Play" from a Computer means that Banelings will never be able to beat marines.

Humans are limited to 300APM (maybe 800 APM in short bursts if you've been training for years). Computers can execute... 8,000+ APM easily. I'm talking about normal PCs... not specialized Multi-GPU dozens-of-KiloWatts supercomputer (like AlphaGo).
First Strike +1/+1 and Indestructible.

User avatar
Diadem
Posts: 5652
Joined: Wed Jun 11, 2008 11:03 am UTC
Location: The Netherlands

Re: The Thread To Remind Me We're Living In The Future

Postby Diadem » Wed May 31, 2017 8:42 am UTC

KnightExemplar wrote:
Mutex wrote:
mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.


I find it hard to believe that StarCraft is a frontier. If an AI is allowed to execute at maximum APM, then it would utterly destroy humans in basically every setting.

A huge part of StarCraft is about Micro, and computers will always execute Micro better than a Human. The main reason why AIs aren't programmed to take advantage of huge APM-tricks is because they're not fun to play against. For example, "Perfect Play" from a Computer means that Banelings will never be able to beat marines.

Humans are limited to 300APM (maybe 800 APM in short bursts if you've been training for years). Computers can execute... 8,000+ APM easily. I'm talking about normal PCs... not specialized Multi-GPU dozens-of-KiloWatts supercomputer (like AlphaGo).

One problem is that there's no clear definition of what "Playing Starcraft" means for an AI. An AI interfaces directly with the computer, with no physical interface in between, allowing for much faster receiving and sending of information. That's a huge benefit for the AI that has nothing to do with how good a player it is. I'm a pretty mediocre starcraft player, but if I had a well working brain interface with my pc, I'm sure I could beat most top players with ease (Well ok that may be an overstatement. But I'd certainly be a much, much, stronger). You're saying that an AI could get 8000+ APM easily, but is that even theoretically possible with a physical keyboard?

So build me an AI that can beat a human player while interfacing to the game via screen and keyboard. Only then will we truly be able to say that AI beats humans at starcraft.

(And then, of course, to make things *fun* for human opponents, we want to limit the AI to some lower APM count. And AI that could still play decently while limited to say "50 APM average, 100 APM peak" would be much more fun to play against than an AI that beats you with a combination of poor strategy and superhuman micro. It's interesting that for starcraft making an AI that can beat average players in an satisfactory manner is actually harder than just making an AI that can beat humans).
It's one of those irregular verbs, isn't it? I have an independent mind, you are an eccentric, he is round the twist
- Bernard Woolley in Yes, Prime Minister

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Wed May 31, 2017 9:24 am UTC

I think you're both over-estimating how important micro is when playing against a human player - perfect micro won't help you when all your units can't shoot up and are getting slaughtered by mutalisks, and a good human player will counter whatever the AI builds. An AI that can macro well and counter the human players builds when dealing with imperfect information, and actually plan an attack that has a chance of success against a target that makes strategic sense, would be amazing to see.

I agree that an AI could totally wipe the floor with a human at micro, but until an AI can come close to a human at strategy, being able to find out what the other player is planning, and being able to deceive the other player as to your intentions, I don't think micro is going to matter much.

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Wed May 31, 2017 5:10 pm UTC

Mutex wrote:I think you're both over-estimating how important micro is when playing against a human player - perfect micro won't help you when all your units can't shoot up and are getting slaughtered by mutalisks, and a good human player will counter whatever the AI builds. An AI that can macro well and counter the human players builds when dealing with imperfect information, and actually plan an attack that has a chance of success against a target that makes strategic sense, would be amazing to see.


Lets say that an AI just always goes M&Ms (Marines and Marauders) vs a Zerg Human player. What will the Zerg Player do?

Again, AIs with perfect Marine Micro will annihilate Banelings. The video I posted earlier was 20 Marines destroying 40-banelings, taking only Stim-pack damage (the Marines never got hit). I'm not sure if you can counter perfectly-micro'd mass-marines, at least in Terrain vs Zerg.

M&Ms can be deployed before Mutalisks or other flying units hit the scene. So the Zerg Player is really only looking at Zerglings, Banelings, and Roaches as their force. Under normal (human-level) situations, Banelings beat Marines. But a perfect AI warps the game and now the Zerg player has no way to beat the early game M&Ms.

-------------

Its much like Chess for Humans vs Computers. Humans have an ability to see extremely far into the future. When you're talking with very good chess players, they talk about how pawn formations can become a problem in the endgame. (30+ moves away). The Computers are essentially blind to that level of depth, most AIs are only ~20 ply deep or so.

But Computers are perfect at executing the next 20 ply or so, or at least damn near perfect. So the "tactical" edge of computers wipe out any "strategic" flaws in the AI.

Perhaps a "deep seeing, strategic" Chess AI will be developed one day. But there's no point really. Chess is mostly about tactics within the next 10 moves. Similarly, StarCraft is innately a game where "proper-micro" grossly determines the results of combat. Perfect-Micro can (and does!) negate strategic problems, much like how Slayer Boxer's immortal Marines beat Lurkers in Brood War.

Diadem wrote:
KnightExemplar wrote:
Mutex wrote:
mfb wrote:We are nearly done. Starcraft is the only game where humans still have a chance (apart from the last category).

I'd be fascinated to see strides made in Starcraft AI. The fog of war means the AI has to operate with imperfect information.


I find it hard to believe that StarCraft is a frontier. If an AI is allowed to execute at maximum APM, then it would utterly destroy humans in basically every setting.

A huge part of StarCraft is about Micro, and computers will always execute Micro better than a Human. The main reason why AIs aren't programmed to take advantage of huge APM-tricks is because they're not fun to play against. For example, "Perfect Play" from a Computer means that Banelings will never be able to beat marines.

Humans are limited to 300APM (maybe 800 APM in short bursts if you've been training for years). Computers can execute... 8,000+ APM easily. I'm talking about normal PCs... not specialized Multi-GPU dozens-of-KiloWatts supercomputer (like AlphaGo).

One problem is that there's no clear definition of what "Playing Starcraft" means for an AI. An AI interfaces directly with the computer, with no physical interface in between, allowing for much faster receiving and sending of information. That's a huge benefit for the AI that has nothing to do with how good a player it is. I'm a pretty mediocre starcraft player, but if I had a well working brain interface with my pc, I'm sure I could beat most top players with ease (Well ok that may be an overstatement. But I'd certainly be a much, much, stronger). You're saying that an AI could get 8000+ APM easily, but is that even theoretically possible with a physical keyboard?

So build me an AI that can beat a human player while interfacing to the game via screen and keyboard. Only then will we truly be able to say that AI beats humans at starcraft.

(And then, of course, to make things *fun* for human opponents, we want to limit the AI to some lower APM count. And AI that could still play decently while limited to say "50 APM average, 100 APM peak" would be much more fun to play against than an AI that beats you with a combination of poor strategy and superhuman micro. It's interesting that for starcraft making an AI that can beat average players in an satisfactory manner is actually harder than just making an AI that can beat humans).


This is true in basically any "twitch" game. A lot of the games that we humans play are limited by our reaction times. A perfect player / AI in a fighting game could just wait until we push a button, and then take advantage of the locked state of the player.

Humans have the best chance at strategic, long-term thinking games. Poker, Chess, Go, etc. etc. Even if we were to make an artificial limitation of APM for computers mapped to a physical keyboard... that problem is basically already solved.

Machines have always outperformed humans at mechanical tasks of precision. A pick-and-place machine will more accurately place electronic components than a human soldering expert.

So even if we forced the computer / AI to work "through a keyboard", that's only a mechanical movement that has long-been perfected in the industry.
Last edited by KnightExemplar on Wed May 31, 2017 5:22 pm UTC, edited 1 time in total.
First Strike +1/+1 and Indestructible.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Wed May 31, 2017 5:20 pm UTC

KnightExemplar wrote:Lets say that an AI just always goes M&Ms (Marines and Marauders) vs a Zerg Human player. What will the Zerg Player do?

I'm not much of a SC player, but how about playing defensively with queens and spine crawlers, and macroing until you have an army you can surprise-drop into the AI base while its forces are halfway across the map because it didn't detect you sneaking past them. No matter how perfect its micro is, if it can't do strategy or work out what the other player is doing, good SC players will very quickly find a way to beat it every time.

Don't get it wrong, perfect micro is a hell of an advantage and arguably if the AI got to the point of being reasonably decent at strategy, its micro would be enough to make it win every time. But even getting an AI to that point in SC strategy would be fascinating, given the challenges the game poses compared to turn-based board games where you can see the entire board.

User avatar
sardia
Posts: 5843
Joined: Sat Apr 03, 2010 3:39 am UTC

Re: The Thread To Remind Me We're Living In The Future

Postby sardia » Wed May 31, 2017 5:38 pm UTC

Mutex wrote:
KnightExemplar wrote:Lets say that an AI just always goes M&Ms (Marines and Marauders) vs a Zerg Human player. What will the Zerg Player do?

I'm not much of a SC player, but how about playing defensively with queens and spine crawlers, and macroing until you have an army you can surprise-drop into the AI base while its forces are halfway across the map because it didn't detect you sneaking past them. No matter how perfect its micro is, if it can't do strategy or work out what the other player is doing, good SC players will very quickly find a way to beat it every time.

Don't get it wrong, perfect micro is a hell of an advantage and arguably if the AI got to the point of being reasonably decent at strategy, its micro would be enough to make it win every time. But even getting an AI to that point in SC strategy would be fascinating, given the challenges the game poses compared to turn-based board games where you can see the entire board.

Alright, you caught up to my original post. What they did as a workaround was to attack first and fast. The ai can rush faster than humans because of superior micro. So you program it to have three good rush strategies, which forces the human to react or worse, guess what the counter is. If humans can't, it's an easy win for the ai.
This programming has the advantage of not relying on the infinite possibility by condensing the game down to the early game and attacking first.

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Wed May 31, 2017 5:51 pm UTC

Mutex wrote:
KnightExemplar wrote:Lets say that an AI just always goes M&Ms (Marines and Marauders) vs a Zerg Human player. What will the Zerg Player do?

I'm not much of a SC player, but how about playing defensively with queens and spine crawlers, and macroing until you have an army you can surprise-drop into the AI base while its forces are halfway across the map because it didn't detect you sneaking past them. No matter how perfect its micro is, if it can't do strategy or work out what the other player is doing, good SC players will very quickly find a way to beat it every time.

Don't get it wrong, perfect micro is a hell of an advantage and arguably if the AI got to the point of being reasonably decent at strategy, its micro would be enough to make it win every time. But even getting an AI to that point in SC strategy would be fascinating, given the challenges the game poses compared to turn-based board games where you can see the entire board.


Are you aware of the Poker AI that beat a group of humans at Texas Holdem?

Poker is a far better game of "incomplete information" than StarCraft, in that Poker is purely about incomplete information... and doesn't have little things like "Micro" that could potentially dominate the game. Winning in a game of incomplete information uses an algorithm called "Counterfactual Regret Minimization" which basically looks for a Nash-Equilibrium like structure in the decision space.

In effect, it means that the "guesses" the computer makes are perfectly randomized (or at least, as well randomized as our randomization algorithms get). And the computer plays against itself to look for Nash Equalibriums (or something like them) ahead of time. IIRC, the Texas Hold'em Bot played Poker against itself for ~6 months, and then continuously played Poker against itself and simulations of its humans opponents each night of the ~30 day Poker Match (continuously improving during that whole time).

Just as you can't beat a Dice in "Rock Paper Scissors" (Rock on 1/2, Paper on 3/4, Scissors on 5/6), you can't beat an AI that has discovered the "equilibrium" point of any game of incomplete information. You can at best tie the AI's randomization strategy (ie: Defined as a win 50% of the time).

----------

In the case of Poker, the randomization strategy is the probability the CPU should "bluff" (pretend it has a good hand), or "fold". Now that its playing at close to a Nash Equilibrium, human players can't tell if the AI is bluffing anymore... and the AI begins to win in Poker.
First Strike +1/+1 and Indestructible.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Wed May 31, 2017 6:12 pm UTC

Can you scout the other player's hand in Poker?

Because that's the other aspect of SC, the need to try to obtain knowledge about the other player by scouting their base. Not to mention occasionally checking around the map for any hidden bases, and balancing the need to do this with the resources used doing so (losing scouts when they enter the base and get destroyed, but then again if you know where their defensive structures are you might be able to avoid them and get more info...). SC requires competences on many different fronts at once: Taking expansions, building economy, building production, building army, scouting, determining where and when to attack the enemy base, knowing when and where to engage their army, engaging the army (that's where micro comes in!), choosing a tech tree and knowing when to switch in retaliation to the enemy's tech choices, knowing when to research upgrades. That's what makes it such an interesting AI problem to me - there's so much to be competent at, and balance and trade off between.

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Wed May 31, 2017 6:17 pm UTC

Mutex wrote:Can you scout the other player's hand in Poker?

Because that's the other aspect of SC, the need to try to obtain knowledge about the other player by scouting their base. Not to mention occasionally checking around the map for any hidden bases, and balancing the need to do this with the resources used doing so (losing scouts when they enter the base and get destroyed, but then again if you know where their defensive structures are you might be able to avoid them and get more info...). SC requires competences on many different fronts at once: Taking expansions, building economy, building production, building army, scouting, determining where and when to attack the enemy base, knowing when and where to engage their army, engaging the army (that's where micro comes in!), choosing a tech tree and knowing when to switch in retaliation to the enemy's tech choices, knowing when to research upgrades. That's what makes it such an interesting AI problem to me - there's so much to be competent at, and balance and trade off between.


Yeah, and that's what Chess was like before Stockfish 7. You "feel" your opponent out, figure out their emotional state, etc. etc. It turned out that none of that mattered however.

It turns out that you can win at Chess by simply playing perfectly for the next 20 ply / 10-turns ahead with perfect accuracy. Everything else is an implementation detail. Similarly, when your 20 Marines (backed by 8000+ APM) are countering 40 Banelings without taking a single point of damage, I have my doubts that anything else matters in StarCraft.

A huge part of StarCraft is the twitch-play aspect. True, there's an element of strategic play and counter-thinking, but that's built on top of the battles that occur. If an 8000 APM bot can win every battle on shear APM alone, then it will dominate play without much regard to strategy or counterplay. Just as Chess-bots are all about playing dumb heuristics perfectly for the next 10-moves (often completely blind to the super-long term strategic elements of Chess), StarCraft Bots will likely be about playing dumb-heuristics at 8000+ APM to dominate the short-term element of the game... with slight emphasis on the strategic element (maybe something as dumb as "Randomly pick between three Rush Strategies", as Sardia points out)

Why scout when you're faster than the opponent? Why scout when your M&Ms beat everything except a purely defensive strategy (ie: Queens and Spine Crawlers). Just put M&Ms in front of the opponent's base, so that they can't expand... then you expand and overwhelm them with 2 bases or 3-bases.
First Strike +1/+1 and Indestructible.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Wed May 31, 2017 6:25 pm UTC

KnightExemplar wrote:
Mutex wrote:Can you scout the other player's hand in Poker?

Because that's the other aspect of SC, the need to try to obtain knowledge about the other player by scouting their base. Not to mention occasionally checking around the map for any hidden bases, and balancing the need to do this with the resources used doing so (losing scouts when they enter the base and get destroyed, but then again if you know where their defensive structures are you might be able to avoid them and get more info...). SC requires competences on many different fronts at once: Taking expansions, building economy, building production, building army, scouting, determining where and when to attack the enemy base, knowing when and where to engage their army, engaging the army (that's where micro comes in!), choosing a tech tree and knowing when to switch in retaliation to the enemy's tech choices, knowing when to research upgrades. That's what makes it such an interesting AI problem to me - there's so much to be competent at, and balance and trade off between.


Yeah, and that's what Chess was like before Stockfish 7. You "feel" your opponent out, figure out their emotional state, etc. etc. It turned out that none of that mattered however.

It turns out that you can win at Chess by simply playing perfectly for the next 20 ply / 10-turns ahead with perfect accuracy. Everything else is an implementation detail. Similarly, when your 20 Marines (backed by 8000+ APM) are countering 40 Banelings without taking a single point of damage, I have my doubts that anything else matters in StarCraft.

A huge part of StarCraft is the twitch-play aspect. True, there's an element of strategic play and counter-thinking, but that's built on top of the battles that occur. If an 8000 APM bot can win every battle on shear APM alone, then it will dominate play without much regard to strategy or counterplay.

While I understand your thinking, I don't think 20 ply and micro are close enough to make micro the only thing that matters.

The best Chess players can think 30 ply ahead, while the AI can do 20 ply perfectly - literally comparing every possible path.

A good SC player is thinking several minutes ahead, micro is just the next couple of seconds.

The Chess AI shows that only being able to think 2/3 as far ahead as your opponent doesn't matter if you can literally evaluate every single possible move. Micro isn't close to 2/3 as far as an SC player can plan.
Last edited by Mutex on Wed May 31, 2017 6:36 pm UTC, edited 1 time in total.

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Wed May 31, 2017 6:35 pm UTC

Mutex wrote:While I understand your thinking, I don't think 20 ply and micro are close enough to make micro the only thing that matters.

The best chess players can think 30 ply ahead, while the AI can do 20 ply perfectly - literally comparing every possible path.

A good SC player is thinking several minutes ahead, micro is just the next couple of seconds.


I'm not sure if you fully comprehend how 8000 APM changes the game of StarCraft.

At 8000 APM, Zerglings beat Siege Tanks and Marines beat +Speed Banelings.

The game is warped, and upside down at that point. It makes no damn sense. 8000 APM destroys our current understanding of strategic play in StarCraft. When Zerglings are murdering Siege Tanks, the "strategic" elements of StarCraft start to break down. All that "thinking ahead" and "building counters" means nothing when pure Micro can warp the "counterplay" of the game.
Last edited by KnightExemplar on Wed May 31, 2017 6:40 pm UTC, edited 1 time in total.
First Strike +1/+1 and Indestructible.

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Wed May 31, 2017 6:39 pm UTC

Yeah, I get it. Although you don't even need high APM to do that with zerglings against tanks. Just spread them out loads and send them to different tanks. Siege tanks are pretty bad against swarms usually. A decent SC player will have some marines in front of their tanks to stop them getting murdered by swarms like that.
Last edited by Mutex on Wed May 31, 2017 7:07 pm UTC, edited 1 time in total.

User avatar
SDK
Posts: 556
Joined: Thu May 22, 2014 7:40 pm UTC
Location: Canada

Re: The Thread To Remind Me We're Living In The Future

Postby SDK » Wed May 31, 2017 7:02 pm UTC

KnightExemplar wrote:So even if we forced the computer / AI to work "through a keyboard", that's only a mechanical movement that has long-been perfected in the industry.

You missed the more important "screen" part of that quote. If the machine has to actually watch a physical screen and type in commands on a physical keyboard, it will be severely hampered. The AI revolution isn't going to happen quite yet.
The biggest number (63 quintillion googols in debt)

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Wed May 31, 2017 8:01 pm UTC

SDK wrote:
KnightExemplar wrote:So even if we forced the computer / AI to work "through a keyboard", that's only a mechanical movement that has long-been perfected in the industry.

You missed the more important "screen" part of that quote. If the machine has to actually watch a physical screen and type in commands on a physical keyboard, it will be severely hampered. The AI revolution isn't going to happen quite yet.


While it'd be a difficult engineering problem to view things from a computer screen... computers are superior to humans at carefully defined image recognition problems. This isn't the "self-driving" car problem, where you come across new stimuli or anything. We're talking about a game that has pre-defined graphics and a limited scope to what can even happen on the screen.

An in those circumstances, a computer is king. Mostly because of the limited search space. Indeed, my example of "Pick and Place" machines is a perfect example of how high-powered computers can optically recognize a pattern and then accurately place components on a motherboard at an outstanding speed far beyond the capabilities of a human.

---------

Anyway, my point is that a game like StarCraft would not be "interesting" from an AI perspective. Go, Poker, Chess... these games were far better as they were more pure tests of intelligence. StarCraft has a huge muscle-memory aspect to it. Just as I'd expect a "Mario Kart" AI to perform "Snaking" (in older games) or "Fire Hopping" (in MK8 non-Deluxe) better than a human, I'd expect an 8000-APM AI to perform "Micro" far far better than a human.

Even if said AI were forced to use a keyboard and mouse, and look at a screen. Each of those problems has been individually solved already. I mean hell, Tesla's "Autopilot" feature is mostly camera-based with image processing (and that's a far, far harder problem). So no, I don't find the "image recognition" or "keyboard pushing" parts of challenge interesting.

Just like we don't expect Chess AIs to perform image-processing to figure out the state of the Chess Board... we don't really expect a StarCraft AI to perform image processing to figure out the state of the StarCraft Map.

-------------

Lets look at a different game. Do you think you can beat an AI-controlled Sumo-bot with an RC Sumobot of your own? Computers have always perform at speeds that far exceed human reaction. That's not AI, that's just a machine performing simple movements far faster than a human can possibly process.
First Strike +1/+1 and Indestructible.

commodorejohn
Posts: 961
Joined: Thu Dec 10, 2009 6:21 pm UTC
Location: Placerville, CA
Contact:

Re: The Thread To Remind Me We're Living In The Future

Postby commodorejohn » Wed May 31, 2017 11:56 pm UTC

I've said it before, but I won't be really impressed until AIs can solve a Roberta Williams adventure game with appreciably less trouble than a human being.
"'Legacy code' often differs from its suggested alternative by actually working and scaling."
- Bjarne Stroustrup
www.commodorejohn.com - in case you were wondering, which you probably weren't.

mfb
Posts: 944
Joined: Thu Jan 08, 2009 7:48 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby mfb » Thu Jun 01, 2017 1:31 am UTC

KnightExemplar wrote:Just as you can't beat a Dice in "Rock Paper Scissors" (Rock on 1/2, Paper on 3/4, Scissors on 5/6), you can't beat an AI that has discovered the "equilibrium" point of any game of incomplete information. You can at best tie the AI's randomization strategy (ie: Defined as a win 50% of the time).

----------

In the case of Poker, the randomization strategy is the probability the CPU should "bluff" (pretend it has a good hand), or "fold". Now that its playing at close to a Nash Equilibrium, human players can't tell if the AI is bluffing anymore... and the AI begins to win in Poker.
You can't beat a die, but the die can't beat you either.
The Poker AI didn't break even. It won by large margin. It exploited weaknesses of the human players.
It turns out that you can win at Chess by simply playing perfectly for the next 20 ply / 10-turns ahead with perfect accuracy. Everything else is an implementation detail.
All good computer chess programs use heuristics to evaluate the value of the board state after x moves, and the quality of the program depends on what you dismissed as "implementation detail".

KnightExemplar
Posts: 5489
Joined: Sun Dec 26, 2010 1:58 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby KnightExemplar » Thu Jun 01, 2017 4:11 am UTC

mfb wrote:
KnightExemplar wrote:Just as you can't beat a Dice in "Rock Paper Scissors" (Rock on 1/2, Paper on 3/4, Scissors on 5/6), you can't beat an AI that has discovered the "equilibrium" point of any game of incomplete information. You can at best tie the AI's randomization strategy (ie: Defined as a win 50% of the time).

----------

In the case of Poker, the randomization strategy is the probability the CPU should "bluff" (pretend it has a good hand), or "fold". Now that its playing at close to a Nash Equilibrium, human players can't tell if the AI is bluffing anymore... and the AI begins to win in Poker.
You can't beat a die, but the die can't beat you either.
The Poker AI didn't break even. It won by large margin. It exploited weaknesses of the human players.


Indeed. In more complicated games, you have to play perfectly to tie if the opponent is playing the Nash Equalibirum. Its harder to see in Rock Paper Scissors because the game is so simple. But the concept is there.

Of course, Poker's Nash Equalibirum hasn't been "solved" yet. But the Poker AI is far closer to it than the humans are. Furthermore, with decades of RNG-research (as well as equipment that can lead to perfect randomness: like Heat-noise RNGs or whatnot), computers can achieve a far more "perfectly random" play, making it hard or impossible to fully predict the bluff behavior.

For details read this. IIRC, Libratus (the winning Poker-AI last February) didn't quite go for "Nash Equalibirum" but computed something very similar (Probably something easier to compute, but maybe not as "strong"). But the paper gives a good introduction. Playing the Nash Equilibrium works. The hard part (the "AI" part anyway...) is computing it. It has already been mathematically proven that Nash Equilibrium exist in every game (where "Game" is defined in Game Theory)

It turns out that you can win at Chess by simply playing perfectly for the next 20 ply / 10-turns ahead with perfect accuracy. Everything else is an implementation detail.
All good computer chess programs use heuristics to evaluate the value of the board state after x moves, and the quality of the program depends on what you dismissed as "implementation detail".


And the quality of said program is simply various degrees of "far better than humans", especially if combined with a Tablebase.

Computers have solved 6-man Chess. If there are 6 or fewer pieces on the board, any computer can play perfectly. This combined with a (even shitty) forward looking heuristic that can play perfectly for ~20-ply or so leads to super-human chess play.
First Strike +1/+1 and Indestructible.

User avatar
Zamfir
I built a novelty castle, the irony was lost on some.
Posts: 7312
Joined: Wed Aug 27, 2008 2:43 pm UTC
Location: Nederland

Re: The Thread To Remind Me We're Living In The Future

Postby Zamfir » Thu Jun 01, 2017 1:08 pm UTC

Anyway, my point is that a game like StarCraft would not be "interesting" from an AI perspective.

Still, researchers are interested in it. Not only in making a crazy fast micro-management miracle, but creating a machine that plays the game more or less as a human would.

The attraction is that Starcraft is much "messier" than abstracted games like chess or poker, and in that sense more like a real-world problem. But all its relevant variables are easily accessible for a computer, and a computer game can obviously be "simulated" much better than real world situations, which avoids costly training data sets.

Edit: I found an article:

https://www.theguardian.com/technology/ ... nd-game-ai

Mutex
Posts: 1064
Joined: Wed Jan 09, 2008 10:32 pm UTC

Re: The Thread To Remind Me We're Living In The Future

Postby Mutex » Thu Jun 01, 2017 1:43 pm UTC

The AI does have some innate advantages, however. One stat that top StarCraft players are ranked on is “actions per minute” (APM): essentially, the amount of times they click each minute. Lacking fingers, muscles, or the possibility of tendonitis, an AI can naturally outclick a human player, which could result in it winning not through strategic thinking but simply by reacting quicker. As a result, DeepMind will be capping the AI at what research scientist Oriol Vinyals describes as “high-level human” speed.

There you go then.

User avatar
Diadem
Posts: 5652
Joined: Wed Jun 11, 2008 11:03 am UTC
Location: The Netherlands

Re: The Thread To Remind Me We're Living In The Future

Postby Diadem » Thu Jun 01, 2017 2:58 pm UTC

KnightExemplar wrote:Machines have always outperformed humans at mechanical tasks of precision. A pick-and-place machine will more accurately place electronic components than a human soldering expert.

So even if we forced the computer / AI to work "through a keyboard", that's only a mechanical movement that has long-been perfected in the industry.

You missed my point. My point wasn't that it is hard for an AI to interact with the game via keyboard and screen. My point was that it would add latency and cap the maximum number of interactions. You simply can't do 8000 apm on a physical keyboard. This means that the AI will be poorer (or rather, slightly less godlike) at micro, meaning it will have to be better at strategy to reach the same level. That's the reason it will be harder to design an AI that can beat humans when the AI has to use a physical keyboard and screen.

Of course if you want to be practical you can just programmatically cap APM, and add some latency. Much easier to design, without the need for special hardware. Which was what I was trying to say in the second part of my post. If you want to write an AI that's not just strong, but also fun to play against, you want to do this anyway.

*edit* Looks like the deepmind team agrees with me. From Zamfir's link:
The end goal for DeepMind is still to build a computer that can play StarCraft the way a person can, by simply looking at the pixels on the screen and sending keyboard and mouse input to manipulate them. But in the short term, it’s a lot easier for DeepMind to teach their system with a simplified view, that spits out low resolution images of the map and mini map, and breaks down the features into different layers, clearly showing detail such as terrain height, unit type, and health.
It's one of those irregular verbs, isn't it? I have an independent mind, you are an eccentric, he is round the twist
- Bernard Woolley in Yes, Prime Minister


Return to “News & Articles”

Who is online

Users browsing this forum: Keyman, sardia and 21 guests