• Log InLog In
  • Register
Liquid`
Team Liquid Liquipedia
EST 13:35
CET 19:35
KST 03:35
  • Home
  • Forum
  • Calendar
  • Streams
  • Liquipedia
  • Features
  • Store
  • EPT
  • TL+
  • StarCraft 2
  • Brood War
  • Smash
  • Heroes
  • Counter-Strike
  • Overwatch
  • Liquibet
  • Fantasy StarCraft
  • TLPD
  • StarCraft 2
  • Brood War
  • Blogs
Forum Sidebar
Events/Features
News
Featured News
Team Liquid Map Contest #2310[ASL22] Finals Preview: Sweet Victory!12[ASL22] Ro4 Preview: Mirror Mirror6[ASL22] Ro8 Preview: Within Reach5[ASL22] Ro8 Preview: In A Tizzy11
Community News
Weekly Cups (Nov 9-15): Clem and herO take over4Dark completes military service, returns to SC217[BSL23] RO16 Group Stage20Weekly Cups (Nov 2-8): herO back on top?2Weekly Cups (Oct 26 - Nov 1): MaxPax, herO, Solar win0
StarCraft 2
General
No more SC2 streamers on team liquid? Team Liquid Map Contest #23 Weekly Cups (Nov 9-15): Clem and herO take over Dark completes military service, returns to SC2 StarCraft open world shooter announced at BlizzCon
Tourneys
soO & FanTaSy return! SC2 Kevin Scrim 4v4 (Nov 16) Sparkling Tuna Cup - Weekly Open Tournament RSL goes to London! 2026 Offline Finals Nov 21-22 Sea Duckling Open (Global, Bronze-Diamond) Dark returns! GSL CK #7 4v4 Team Battle (Nov 19)
Strategy
[H] ZvP Mid-Late Game: Stalkers Collossi HT
Custom Maps
[G] How to playtest arcade maps Starcraft 2 Advance [D] the control group Simple Questions/Answers
External Content
Mutation # 552 Secret Storm The PondCast: SC2 News & Results Mutation # 551 Robotic Revival Mutation # 550 Graveyard Shift
Brood War
General
Light introduces ShieldBattery on his YouTube List of top players active on ShieldBattery BW General Discussion BGH Auto Balance -> http://bghmmr.eu/ NaDa's Body
Tourneys
SC1 Kevin Scrim hosted by SOOP (Nov 18-19) [Megathread] Daily Proleagues [BSL23] Ro16 Group D - Sunday 21:00 CET [BSL23] Ro16 Group B - Saturday 21:00 CET
Strategy
Pluto's TvP 5Marine 2Medic Push Simple Questions, Simple Answers Infested Terran strategies in 2026 Nuke strategy worked in K-League 2026-11-09
Other Games
General Games
The PlayStation 5 Stormgate/Frost Giant Megathread Warcraft III: The Frozen Throne Nintendo Switch Thread Free Space RTS Game, optimised f. Phone/Tablet etc
Dota 2
League of Legends
[TL LoL EUW IHs] Teemo shall perish
Heroes of the Storm
Heroes of the Storm 2.0
Hearthstone
Deck construction bug
TL Mafia
TL Mafia Community Thread
Community
General
US Politics Mega-thread Artificial Intelligence Thread The Big Programming Thread Trading/Investing Thread Russo-Ukrainian War Thread
Fan Clubs
Fan Club Resource Thread Reynor Fan Club The Flash Fan Club
Media & Entertainment
Movie Discussion! Anime Discussion Thread [Manga] One Piece
Sports
Football (Soccer) Thread TeamLiquid Health and Fitness Initiative For 2023
World Cup 2022
Tech Support
Computer Build, Upgrade & Buying Resource Thread
TL Community
The Automated Ban List Recent Gifted Posts
Blogs
The time I organised my…
AssyrianKing
Why I built Repalyzer (1.1.…
BlueStar
Innovate or Game Over?
TrAiDoS
ShieldBattery on Mac
Kraekkling
Can Bots Beat Pros?? Starcr…
namkraft
James Bond movies ranking - pa…
Topin
StarCraft improvement
iopq
Customize Sidebar...

Website Feedback

Closed Threads



Active: 15535 users

BoxeR: "AlphaGo won't beat humans in StarCraft" - Page 16

Forum Index > SC2 General
568 CommentsPost a Reply
Prev 1 14 15 16 17 18 29 Next All
rockslave
Profile Blog Joined January 2008
Brazil318 Posts
March 15 2016 12:00 GMT
#301
On March 15 2016 16:56 Liquid`Bunny wrote:
Show nested quote +
On March 14 2016 21:01 BeyondCtrL wrote:
On March 14 2016 20:51 Liquid`Bunny wrote:
On March 14 2016 20:03 Liquid`Nazgul wrote:
On March 14 2016 19:59 Liquid`Bunny wrote:
Well of course the AI will be able to beat human starcraft players, regardless of it being apm capped or not, as long as they put enough effort into making it. However it would be boring if players didn't take on the challenge of beating it, i myself would love to experience playing against it, we might learn something!

Also it's kind of funny how everyone is viewing the AI winning as humans "losing" I think it would be a great achievement for humanity to make an AI that can learn such a complex task

As long as they create some laws restricting AI from taking over the world~~

When we create a program that can make a better program on it's own, that's when the trouble starts.


AlphaGo does that already.

AlphaGo doesn't change the way it's programmed, it will always be programmed in a certain way. What it can change is parameters within functions to achieve a better result.


Just like humans. You only have about 3 billion variables to program a human, and the great majority of that is about synthesizing proteins and throwing little molecules around.

Which reminds me of a story about Claude Shannon, one of the founders of computer science:
Reporter: Can computers think?
Shannon: Can you think?
Reporter: Yes.
Shannon: So, yes.
What qxc said.
JimmyJRaynor
Profile Blog Joined April 2010
Canada18074 Posts
March 15 2016 12:24 GMT
#302
On March 15 2016 19:24 sh1RoKen wrote:
Because this beast can learn.
And he can do it really fast. He can play 10 million games per day vs himself to understand how the game works and what move or tactic or strategy is better at any possible game position. He can analyse all pro replays in the internet to study how humans play and what is better to use against them. It might sound really hard to believe but that is how it works. He doesn't just calculate all possible moves and chose the best. Google AI thinks much more like human does than you think. And he does it better.

the stuff it will learn playing against itself won't teach it the kinds of tactics that can defeat a human. it might be good for creating its own heuristic functions.

do they plan on making this work without heuristics?
or are you claiming the Google AI will create its own as it studies replays and plays itself?
Ray Kassar To David Crane : "you're no more important to Atari than the factory workers assembling the cartridges"
necaremus
Profile Joined December 2013
45 Posts
Last Edited: 2016-03-15 12:45:37
March 15 2016 12:28 GMT
#303
On March 15 2016 16:59 deacon.frost wrote:
Show nested quote +
On March 15 2016 12:25 necaremus wrote:
i am really interested in this topic as well, but i would side with the "humans would win" opinion

Go is a game with "full information", while sc2 is not. AI can't handle this situation.
if you make it a bo5 with random maps (both players don't know the map beforehand) i doubt the current status of the AI could even beat a mid-class master player.

That's not necessary. This isn't solving by the traditional method(from A you can go to A1 - A56464984651894) but simulating the learning process of the brain. So the biggest obstacle is to transform all the information in the process to the computer so it understands it. Go is much easier for such translation(you have only X-Y, no ramps, not reachable terrain, blank spaces, bases etc.). But the learning itself works the same as our brain.

In SC2 we have multiple high end replay packs so the PC can learn from the best(not sure about BW).

If they do the job properly and PC can play SC without any problems... then the human person will have really tough enemy. Because then it's all about the time and the learning process. And PC can train 24/7


my point about "full information": in SC you have fog of war, you do not see your enemy. There is a big uncertain factor for the AI: do i move out with my army and risk being counter attacked? For the human these factors of uncertainty are normal: to be honest, it's the only way of how we interact with the outside world: with a big uncertainty - we don't even know it any different except for special cases like the game of "Go".

I could imagine that the AI has a big problem, if the AI-scout doesn't find the enemy army (because our human didn't build one, maybe?) it would try to scout the whole map before moving out, because it doesn't want to risk a counter-atk. A human would just a-move and win

On March 15 2016 18:07 Caihead wrote:
As to whether AI's will eventually beat humans at any specific task, in my opinion it's not about whether or not one system has inherent superiority over the other or not, but the amount of energy and resources required to devote to the task to compete.

this.

when i heard about how AlphaGo works, i thought of this: numberphile, knots/DNA
more precise (2nd video to this topic on numberphile ~ 1:30)
Type II Topoisomerase

my thought was "holy shit, we can already 'build' the logical infrastructure, of a component of bacteria"
only thing is: we need about 10^100000 more energy... (arbitrary chosen, but something rly huge as factor)

so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)
“Never assume malice when stupidity will suffice.”
NukeD
Profile Joined October 2010
Croatia1612 Posts
March 15 2016 12:36 GMT
#304
Machines are our friends, not our enemies!
sorry for dem one liners
deacon.frost
Profile Joined February 2013
Czech Republic12134 Posts
March 15 2016 13:25 GMT
#305
On March 15 2016 21:28 necaremus wrote:
Show nested quote +
On March 15 2016 16:59 deacon.frost wrote:
On March 15 2016 12:25 necaremus wrote:
i am really interested in this topic as well, but i would side with the "humans would win" opinion

Go is a game with "full information", while sc2 is not. AI can't handle this situation.
if you make it a bo5 with random maps (both players don't know the map beforehand) i doubt the current status of the AI could even beat a mid-class master player.

That's not necessary. This isn't solving by the traditional method(from A you can go to A1 - A56464984651894) but simulating the learning process of the brain. So the biggest obstacle is to transform all the information in the process to the computer so it understands it. Go is much easier for such translation(you have only X-Y, no ramps, not reachable terrain, blank spaces, bases etc.). But the learning itself works the same as our brain.

In SC2 we have multiple high end replay packs so the PC can learn from the best(not sure about BW).

If they do the job properly and PC can play SC without any problems... then the human person will have really tough enemy. Because then it's all about the time and the learning process. And PC can train 24/7


my point about "full information": in SC you have fog of war, you do not see your enemy. There is a big uncertain factor for the AI: do i move out with my army and risk being counter attacked? For the human these factors of uncertainty are normal: to be honest, it's the only way of how we interact with the outside world: with a big uncertainty - we don't even know it any different except for special cases like the game of "Go".

I could imagine that the AI has a big problem, if the AI-scout doesn't find the enemy army (because our human didn't build one, maybe?) it would try to scout the whole map before moving out, because it doesn't want to risk a counter-atk. A human would just a-move and win

Show nested quote +
On March 15 2016 18:07 Caihead wrote:
As to whether AI's will eventually beat humans at any specific task, in my opinion it's not about whether or not one system has inherent superiority over the other or not, but the amount of energy and resources required to devote to the task to compete.

this.

when i heard about how AlphaGo works, i thought of this: numberphile, knots/DNA
more precise (2nd video to this topic on numberphile ~ 1:30)
Type II Topoisomerase

my thought was "holy shit, we can already 'build' the logical infrastructure, of a component of bacteria"
only thing is: we need about 10^100000 more energy... (arbitrary chosen, but something rly huge as factor)

so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)

It doesn't.

It will analyze probably all the top games, then it will play "some" and analyze these. Now we skip a plenty of tech data. In the end it has an information, that if it has this level of uncertainty, this big army and the enemy is doing that, then it's better to move out than not to. And in the end it could be a wrong move, because of the fact it doesn't know everything. That's the beauty of learning a neural net. If you prepare proper learning scenarios it has similar decision making as a human being. The better learning material the better results.

The "AI" doesn't need ALL the information if the learning models work with that. But to build such "AI" you need time, money, proper learning materials and HW. That's why game developers use AI that cheats and then dumb it down. It's easier to write. (or maybe they used dumbed net, who cares)

Imagine a savant who can ONLY play SC and nothing else. That's a result of properly trained net The question can be what will be the input for the "AI", will there be a limitation on the controlling mechanism of it? And other questions asked by the Dagger of Bisu

It is the exactly same thing as the difference between Koreans and Foreigners. Thanks to the fact that Koreans can train selected scenarios multiple hours in a row they have more optimal solution and these solution can be sometimes abused(you know what the expected result is).
It seems to me that either we are talking about something else or you don't know how that works
I imagine France should be able to take this unless Lilbow is busy practicing for Starcraft III. | KadaverBB is my fairy ban mother.
heqat
Profile Joined October 2011
Switzerland96 Posts
March 15 2016 13:39 GMT
#306
On March 15 2016 21:28 necaremus wrote:
so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)


Regarding this subject, in 2007 the world’s total CPU power was approximatively equal to one human brain.

Reference here:

arstechnica.com
sh1RoKen
Profile Joined March 2012
Russian Federation93 Posts
Last Edited: 2016-03-15 13:53:11
March 15 2016 13:41 GMT
#307
On March 15 2016 21:24 JimmyJRaynor wrote:
Show nested quote +
On March 15 2016 19:24 sh1RoKen wrote:
Because this beast can learn.
And he can do it really fast. He can play 10 million games per day vs himself to understand how the game works and what move or tactic or strategy is better at any possible game position. He can analyse all pro replays in the internet to study how humans play and what is better to use against them. It might sound really hard to believe but that is how it works. He doesn't just calculate all possible moves and chose the best. Google AI thinks much more like human does than you think. And he does it better.

the stuff it will learn playing against itself won't teach it the kinds of tactics that can defeat a human. it might be good for creating its own heuristic functions.

do they plan on making this work without heuristics?
or are you claiming the Google AI will create its own as it studies replays and plays itself?


How they did with GO:

1. They showed to AI 30 million GO moves from the internet marked with "Good move" (of a player who won at this very game) or "Bad move" (of a player who lost at this very game).
At this point AplhaGO learned GO rules just by itself. Without any manual algorithms or instructions given by human. It started to recognize and predict what move a human would do when he is trying to win in common game positions. It has got a human "Intuition".

2. Then it started playing vs itself over and over and over again to create a database of moves evaluated by pure calculations % of winning for every move. And it trained it's intuition even more.

3. Then it combines calculations which human can never achieve with an intuition which human don't really understand to somehow compare it level to what AlphaGO achieved.

They will probably do the same thing with Starcraft. After lesson 1 it will start to play like a really good human. After lesson 2 it will defeat anyone without any chances on "strategy" level. I don't even mention mechanical reaction, speed and accuracy level which wasn't even considered during starcraft balance design.
Be polite, be professional, but have a plan to kill everybody you meet.
necaremus
Profile Joined December 2013
45 Posts
March 15 2016 14:25 GMT
#308
On March 15 2016 22:25 deacon.frost wrote:
It doesn't.

It will analyze probably all the top games, then it will play "some" and analyze these. Now we skip a plenty of tech data. In the end it has an information, that if it has this level of uncertainty, this big army and the enemy is doing that, then it's better to move out than not to. And in the end it could be a wrong move, because of the fact it doesn't know everything. That's the beauty of learning a neural net. If you prepare proper learning scenarios it has similar decision making as a human being. The better learning material the better results.

The "AI" doesn't need ALL the information if the learning models work with that. But to build such "AI" you need time, money, proper learning materials and HW. That's why game developers use AI that cheats and then dumb it down. It's easier to write. (or maybe they used dumbed net, who cares)

Imagine a savant who can ONLY play SC and nothing else. That's a result of properly trained net The question can be what will be the input for the "AI", will there be a limitation on the controlling mechanism of it? And other questions asked by the Dagger of Bisu

It is the exactly same thing as the difference between Koreans and Foreigners. Thanks to the fact that Koreans can train selected scenarios multiple hours in a row they have more optimal solution and these solution can be sometimes abused(you know what the expected result is).
It seems to me that either we are talking about something else or you don't know how that works


I don't know how it works completely, but i have a few information and a lot of uncertainty, which i use to evaluate the situation

I know, that the AI doesn't need all information, but i wanted to point out, that we have a whole new range of problems regarding starcraft compared to go.

let's suppose the AI has the "perfect" strategy: this would mean the AI always plays the exact same way -> the human wouldn't have this "uncertainty", because he knows how the AI is going to play and he could craft a strategy, that isn't perfect, but beats the strategy of the AI (for example a doom-drop? i don't know)

you could try to bypass this and give the AI a range of strategies it can choose from. but if you hardcode this into the AI i don't see the point of even trying to build an AI for starcraft. The hardcoded strategy would be human-created making it a "machine+human vs human" match
“Never assume malice when stupidity will suffice.”
sh1RoKen
Profile Joined March 2012
Russian Federation93 Posts
Last Edited: 2016-03-15 14:52:17
March 15 2016 14:46 GMT
#309
On March 15 2016 23:25 necaremus wrote:
Show nested quote +
On March 15 2016 22:25 deacon.frost wrote:
It doesn't.

It will analyze probably all the top games, then it will play "some" and analyze these. Now we skip a plenty of tech data. In the end it has an information, that if it has this level of uncertainty, this big army and the enemy is doing that, then it's better to move out than not to. And in the end it could be a wrong move, because of the fact it doesn't know everything. That's the beauty of learning a neural net. If you prepare proper learning scenarios it has similar decision making as a human being. The better learning material the better results.

The "AI" doesn't need ALL the information if the learning models work with that. But to build such "AI" you need time, money, proper learning materials and HW. That's why game developers use AI that cheats and then dumb it down. It's easier to write. (or maybe they used dumbed net, who cares)

Imagine a savant who can ONLY play SC and nothing else. That's a result of properly trained net The question can be what will be the input for the "AI", will there be a limitation on the controlling mechanism of it? And other questions asked by the Dagger of Bisu

It is the exactly same thing as the difference between Koreans and Foreigners. Thanks to the fact that Koreans can train selected scenarios multiple hours in a row they have more optimal solution and these solution can be sometimes abused(you know what the expected result is).
It seems to me that either we are talking about something else or you don't know how that works


I don't know how it works completely, but i have a few information and a lot of uncertainty, which i use to evaluate the situation

I know, that the AI doesn't need all information, but i wanted to point out, that we have a whole new range of problems regarding starcraft compared to go.

let's suppose the AI has the "perfect" strategy: this would mean the AI always plays the exact same way -> the human wouldn't have this "uncertainty", because he knows how the AI is going to play and he could craft a strategy, that isn't perfect, but beats the strategy of the AI (for example a doom-drop? i don't know)

you could try to bypass this and give the AI a range of strategies it can choose from. but if you hardcode this into the AI i don't see the point of even trying to build an AI for starcraft. The hardcoded strategy would be human-created making it a "machine+human vs human" match


If AI will succeed at finding perfect strategy which will win 100% of times and can't be countered by any counter-action of his opponent, he will execute it over and over again without any chance of loosing. Otherwise that strategy can't be called perfect and AI wouldn't play it over and over again. He knows what predictability is and will vary his buildorders.

He is programmed in a way that he will try to do every move that increases his chance of winning and will avoid any situations that decreases his chance of winning.

He can blink micro for 16 hours if he knows that that will 100% give him 1 HP advantage over his opponent. But he will never go for all-in with 98% chance of winning if there is any possible way to increase it for another 0.000000001%.
Be polite, be professional, but have a plan to kill everybody you meet.
sh1RoKen
Profile Joined March 2012
Russian Federation93 Posts
Last Edited: 2016-03-15 14:51:29
March 15 2016 14:50 GMT
#310
Be polite, be professional, but have a plan to kill everybody you meet.
Pwere
Profile Joined April 2010
Canada1557 Posts
Last Edited: 2016-03-15 14:53:25
March 15 2016 14:52 GMT
#311
Sorry to say this, necaremus, but it seems to me you don't understand how AI works. Some of us have an advanced degree in AI, but since it's not exactly the same branch, we don't feel comfortable making predictions.

What you are describing is absolutely a non-issue for this type of AI. You are thinking of a bot, which is vastly inferior. There is nothing inherently difficult about Starcraft for an AI. The strategic aspect is complex, but you only have to be better than humans, not perfect. And humans waste most of their training regimen on mechanics.

AI these days are at least on par with humans when dealing with uncertainty. Pure numbers over thousands of games beat intuition. They don't even bother with profiling to exploit people's weaknesses because of how dominant the analytical approach is when dealing with uncertainty.

All the "problems" pointed out in this thread are mostly annoyances. Not being able to simulate millions of games per day is the bigger struggle, but I feel comfortable saying you can easily run hundreds of games of Broodwar per hour on gaming hardware, so Google would find a way.

I think the main reason Google even considers Starcraft is that it's fun to watch, and millions of people would watch these games. It would be a publicity stunt.
thePunGun
Profile Blog Joined January 2016
598 Posts
March 15 2016 14:57 GMT
#312
On March 15 2016 22:41 sh1RoKen wrote:
Show nested quote +
On March 15 2016 21:24 JimmyJRaynor wrote:
On March 15 2016 19:24 sh1RoKen wrote:
Because this beast can learn.
And he can do it really fast. He can play 10 million games per day vs himself to understand how the game works and what move or tactic or strategy is better at any possible game position. He can analyse all pro replays in the internet to study how humans play and what is better to use against them. It might sound really hard to believe but that is how it works. He doesn't just calculate all possible moves and chose the best. Google AI thinks much more like human does than you think. And he does it better.

the stuff it will learn playing against itself won't teach it the kinds of tactics that can defeat a human. it might be good for creating its own heuristic functions.

do they plan on making this work without heuristics?
or are you claiming the Google AI will create its own as it studies replays and plays itself?


How they did with GO:

1. They showed to AI 30 million GO moves from the internet marked with "Good move" (of a player who won at this very game) or "Bad move" (of a player who lost at this very game).
At this point AplhaGO learned GO rules just by itself. Without any manual algorithms or instructions given by human. It started to recognize and predict what move a human would do when he is trying to win in common game positions. It has got a human "Intuition".

2. Then it started playing vs itself over and over and over again to create a database of moves evaluated by pure calculations % of winning for every move. And it trained it's intuition even more.

3. Then it combines calculations which human can never achieve with an intuition which human don't really understand to somehow compare it level to what AlphaGO achieved.

They will probably do the same thing with Starcraft. After lesson 1 it will start to play like a really good human. After lesson 2 it will defeat anyone without any chances on "strategy" level. I don't even mention mechanical reaction, speed and accuracy level which wasn't even considered during starcraft balance design.


Well it's not that simple considering, that SC is not turn based. There are much more complex calculations involved. I'm sure it will get there eventually, but it won't learn as fast in a real-time startegy game like SC, since it's way more random.

"You cannot teach a man anything, you can only help him find it within himself."
BaronVonOwn
Profile Joined April 2011
299 Posts
March 15 2016 14:58 GMT
#313
I'm sure BoxeR's only saying this because he'd love the publicity, because any serious AI would make a pro SC2 player look like a bronze noob. AI's have perfect mechanics meaning he will lose every micro/macro battle. Mechanics dominate strategy in SC2 and you can win games based on pure micro/macro alone. Starcraft was developed assuming human players with lag and poor reactions. A lot of the game elements would be rendered useless with an AI player. For example think about raven seeker missiles. Those will basically never hit against a properly coded AI. Hell they never hit against human players either.
sh1RoKen
Profile Joined March 2012
Russian Federation93 Posts
March 15 2016 15:13 GMT
#314
On March 15 2016 23:57 thePunGun wrote:
Show nested quote +
On March 15 2016 22:41 sh1RoKen wrote:
On March 15 2016 21:24 JimmyJRaynor wrote:
On March 15 2016 19:24 sh1RoKen wrote:
Because this beast can learn.
And he can do it really fast. He can play 10 million games per day vs himself to understand how the game works and what move or tactic or strategy is better at any possible game position. He can analyse all pro replays in the internet to study how humans play and what is better to use against them. It might sound really hard to believe but that is how it works. He doesn't just calculate all possible moves and chose the best. Google AI thinks much more like human does than you think. And he does it better.

the stuff it will learn playing against itself won't teach it the kinds of tactics that can defeat a human. it might be good for creating its own heuristic functions.

do they plan on making this work without heuristics?
or are you claiming the Google AI will create its own as it studies replays and plays itself?


How they did with GO:

1. They showed to AI 30 million GO moves from the internet marked with "Good move" (of a player who won at this very game) or "Bad move" (of a player who lost at this very game).
At this point AplhaGO learned GO rules just by itself. Without any manual algorithms or instructions given by human. It started to recognize and predict what move a human would do when he is trying to win in common game positions. It has got a human "Intuition".

2. Then it started playing vs itself over and over and over again to create a database of moves evaluated by pure calculations % of winning for every move. And it trained it's intuition even more.

3. Then it combines calculations which human can never achieve with an intuition which human don't really understand to somehow compare it level to what AlphaGO achieved.

They will probably do the same thing with Starcraft. After lesson 1 it will start to play like a really good human. After lesson 2 it will defeat anyone without any chances on "strategy" level. I don't even mention mechanical reaction, speed and accuracy level which wasn't even considered during starcraft balance design.


Well it's not that simple considering, that SC is not turn based. There are much more complex calculations involved. I'm sure it will get there eventually, but it won't learn as fast in a real-time startegy game like SC, since it's way more random.



There is nothing random in Starcraft. It might be random for humans but for computer it is 100% predictable.

It will definitely take much more time to learn lesson 1 because of the more complicated design than GO.
But Man. Some enthusiast were managed to teach another much more simpler artificial neural network to complete Mario level in 34 attempts! And this was a child play compared to what google is capable of in both intellectual and hardware resources.
Be polite, be professional, but have a plan to kill everybody you meet.
ClanRH.TV
Profile Joined July 2010
United States462 Posts
March 15 2016 15:27 GMT
#315
On March 15 2016 22:39 heqat wrote:
Show nested quote +
On March 15 2016 21:28 necaremus wrote:
so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)


Regarding this subject, in 2007 the world’s total CPU power was approximatively equal to one human brain.

Reference here:

arstechnica.com


That study is saying that the world's total storage capacity is the same as an adult humans DNA, not that the world's CPU power is approximately equal to one human brain. The title that article uses and the actual conclusions are very misleading.
"Don't take life too seriously because you'll never get out alive."
necaremus
Profile Joined December 2013
45 Posts
March 15 2016 15:29 GMT
#316
On March 15 2016 23:52 Pwere wrote:
Sorry to say this, necaremus, but it seems to me you don't understand how AI works. Some of us have an advanced degree in AI, but since it's not exactly the same branch, we don't feel comfortable making predictions.

What you are describing is absolutely a non-issue for this type of AI. You are thinking of a bot, which is vastly inferior. There is nothing inherently difficult about Starcraft for an AI. The strategic aspect is complex, but you only have to be better than humans, not perfect. And humans waste most of their training regimen on mechanics.

AI these days are at least on par with humans when dealing with uncertainty. Pure numbers over thousands of games beat intuition. They don't even bother with profiling to exploit people's weaknesses because of how dominant the analytical approach is when dealing with uncertainty.

i do agree, that we have a different picture of the situation. I also know, that my "hardcoded" example is pointing towards a bot and the AI the deepmind team created is entirely different.

A weakness of mine may very well be, that i do not fear to make predictions, although my range of information is very limited (like in this case). I do make these prediction on the one hand to find out where i might be wrong (because of lack of information) and on the other hand... because it's fun for me

some people think this attitude is annoying, but - as far as i know - it's the fastest way of learning and improving oneself. Maybe people think this kind of thing is annoying, because they make the mistake of interpreting evaluations as facts or opinions.

but i wanna get back to the AI and starcraft... and i wanna try to explain why i may build a different picture as you do

i didn't say it, but when i evaluated the AI vs human thingy, i didn't take the state of the game as is, but a slightly different version of starcraft, where AI and human would be on-par if it comes to micro-management of the units:
both the human and the AI would use the same algorithm for blink-micro, concave building, focus fire (and so on) reducing the game to positional advantage on the map and build-order strategy.

i did this, because you would not need the deepmind AI to win against a human if you take the state of the game as of now.
the superior micro-control of a simple AI (which even i could program) would win pretty much every game against a real human...
“Never assume malice when stupidity will suffice.”
necaremus
Profile Joined December 2013
45 Posts
March 15 2016 15:49 GMT
#317
On March 16 2016 00:27 ClanRH.TV wrote:
Show nested quote +
On March 15 2016 22:39 heqat wrote:
On March 15 2016 21:28 necaremus wrote:
so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)


Regarding this subject, in 2007 the world’s total CPU power was approximatively equal to one human brain.

Reference here:

arstechnica.com


That study is saying that the world's total storage capacity is the same as an adult humans DNA, not that the world's CPU power is approximately equal to one human brain. The title that article uses and the actual conclusions are very misleading.

ty
I actually didn't bother reading it, because i thought "total nonsense" when i saw the title. But your statement suggest, that it may very well be worth a read... just a bad title :3
“Never assume malice when stupidity will suffice.”
heqat
Profile Joined October 2011
Switzerland96 Posts
March 15 2016 16:09 GMT
#318
On March 16 2016 00:49 necaremus wrote:
Show nested quote +
On March 16 2016 00:27 ClanRH.TV wrote:
On March 15 2016 22:39 heqat wrote:
On March 15 2016 21:28 necaremus wrote:
so... before we come and surpass the human... there is a loooooong way ahead of us: maybe we need more energy than our sun can provide in its lifetime to simulate a human. (i'm not saying it's impossible ;p)


Regarding this subject, in 2007 the world’s total CPU power was approximatively equal to one human brain.

Reference here:

arstechnica.com


That study is saying that the world's total storage capacity is the same as an adult humans DNA, not that the world's CPU power is approximately equal to one human brain. The title that article uses and the actual conclusions are very misleading.

ty
I actually didn't bother reading it, because i thought "total nonsense" when i saw the title. But your statement suggest, that it may very well be worth a read... just a bad title :3


Sure, but still from the article:

"To put our findings in perspective, the 6.4*1018 instructions per second that human kind can carry out on its general-purpose computers in 2007 are in the same ballpark area as the maximum number of nerve impulses executed by one human brain per second,"

Of course the brain works very differently than a CPU, so we cannot directly compare them in term of power.

nimdil
Profile Blog Joined January 2011
Poland3775 Posts
March 15 2016 19:40 GMT
#319
It's funny that people discuss - at the same time - if AI could beat top players in StarCraft and the ways how SC-tuned AlphaGo like AI (so AlphaSC I guess) should be obstructed so that the game will be fair.

AlphaGo is an AI that bases it's aactions on graphical input and that's it. If you feel that you need to tune down the ability of AI to execute perfect strategies at superhuman speed even though it would be using standard inputs then sorry but it is game over. AI won.
Xyik
Profile Blog Joined November 2009
Canada728 Posts
March 15 2016 19:45 GMT
#320
For those claiming that it will learn from games on the internet, how many replays / pieces of game data is there available online? I would guess < 100K high-level games (and thats being generous, I would guess even fewer as top pros rarely release replays).

Lets say we have 100,000 replays.

divide that by the 9 possible match-ups (TvX, ZvX, PvX), now we only have 11111 per match-up.
divide that by lets say, 10 popular maps within the last 2 years. Now we have ~1000 replays per map per match-up.
divide that by the number of build order openings / start-positions and we have at the most, 100 replays to study from in each match-up on a particular map on particular openings.

I don't think thats enough data to properly seed the A.I, so most of its learning will be from playing itself which will be quite difficult. Lets say Google uses a cluster of 10,000 machines, each running a copy of SC for the A.I to play, that allows it play maybe 2M games a day against itself. Now do the same division to figure how many games it will be able to play for each map / match-up / starting position / build-order a day.

I don't know how much data AlphaGo needed to reach its current level in Go, but clearly training the A.I in SC will be a much more difficult task based on acquiring enough data alone.

Then there is the challenge of it actually being able to learn the nuances of the game and interpret game-state. Even if Google found a way to sufficiently train it, I am really not convinced it could win.

I think whats really important is map-level data as well. will the A.I be able to interpret what the map looks like?
Prev 1 14 15 16 17 18 29 Next All
Please log in or register to reply.
Live Events Refresh
OSC
01:00
OSC 5000
EmpressLilyy30
Liquipedia
Patches Events
23:00
Patch Clash S1: Monthly #4
RotterdaM224
Liquipedia
[ Submit Event ]
Live Streams
Refresh
StarCraft 2
Serral 1568
ProTech147
MindelVK 38
EmSc Tv 14
elazer 1
StarCraft: Brood War
Britney 17038
Shuttle 574
Dewaltoss 75
firebathero 55
910 17
soO 11
yabsab 8
ivOry 7
Dota 2
qojqva3320
420jenkins365
monkeys_forever276
League of Legends
Trikslyr67
Counter-Strike
pashabiceps1685
Other Games
singsing3032
FrodaN1234
Beastyqt446
B2W.Neo306
RotterdaM224
ArmadaUGS151
QueenE102
mouzStarbuck98
KnowMe69
DeMusliM59
UpATreeSC41
Railgan12
Organizations
Other Games
BasetradeTV147
StarCraft 2
angryscii 24
EmSc Tv 14
EmSc2Tv 14
[ Show 13 non-featured ]
StarCraft 2
• Adnapsc2 18
• Reevou 10
• AfreecaTV YouTube
• intothetv
• Kozan
• IndyKCrew
• Migwel
StarCraft: Brood War
• BSLYoutube
• STPLYoutube
• ZZZeroYoutube
Dota 2
• WagamamaTV674
Other Games
• imaqtpie479
• Shiphtur372
Upcoming Events
Replay Cast
4h 25m
KCM Race Survival
15h 25m
Ultimate Battle
17h 25m
ZerO vs Bisu
RSL Revival
1d 16h
BSL
2 days
eOnzErG vs StRyKeR
Cross vs JDConan
RSL Revival
2 days
IPSL
2 days
Ret vs ggnore13
nOmaD vs Aether
BSL
3 days
Bonyth vs Doodle
Kyrie vs Jimin
Replay Cast
3 days
WardiTV Weekly
3 days
[ Show More ]
Kung Fu Cup
4 days
Solar vs Classic
ByuN vs SHIN
PiGosaur Cup
5 days
The PondCast
6 days
Liquipedia Results

Completed

Proleague 2026-11-17
Blizzard Classic Cup 2026
HCC Season 3

Ongoing

Acropolis #5
KCM Race Survival 2026 Season 4
International ShieldBattery League
BSL Season 23
IPSL Winter 2026-27
Calamity Cup Div. A Season 8
Underdog Cup 2026 Div 2-3
Stake Ranked Episode 6
BLAST Rivals Fall 2026
IEM Beijing 2026
PGL Masters Bucharest 2026
Stake Ranked Episode 5
1win Private Club #2
Thunderpick World Champ. '26
ESL Pro League Season 24
Stake Ranked Episode 4
1win Private Club #1
Logitech G Play Connect 2026
SL StarSeries Fall 2026
FISSURE Playground #3
BLAST Open Fall 2026

Upcoming

Sector 591
YSL S4
SC4ALL II: Brood War
NBSL 2026 Xiamen LAN
BSL 23: Non-Korean Championship
HSC XXX
Stellar Fest 2: Lunar Cup
StarCraft2 Weekly Open Tournament 2026 Grand Finals
SC4ALL II: StarCraft II
Kung Fu Cup 2026 Grand Finals
RSL Offline Finals
IEM Kraków 2027
PGL Bucharest 2027
BLAST Bounty Winter 2027
PGL Major Singapore 2026
TLPD

1. ByuN
2. TY
3. Dark
4. Solar
5. Stats
6. Nerchio
7. sOs
8. soO
9. INnoVation
10. Elazer
1. Rain
2. Flash
3. EffOrt
4. Last
5. Bisu
6. Soulkey
7. Mini
8. Sharp
Sidebar Settings...

Advertising | Privacy Policy | Terms Of Use | Contact Us

Original banner artwork: Jim Warren
The contents of this webpage are copyright © 2026 TLnet. All Rights Reserved.