Humans Are Still Better Than AI at StarCraft (technologyreview.com)

29-year-old professional StarCraft player Song Byung-gu won 4-0 in the world's first contest between AI systems and professional human players, writes MIT Technology Review. An anonymous reader quotes their report: One of the bots, dubbed "CherryPi," was developed by Facebook's AI research lab. The other bots came from Australia, Norway, and Korea. The contest took place at Sejong University in Seoul, Korea, which has hosted annual StarCraft AI competitions since 2010. Those previous events matched AI systems against each other (rather than against humans) and were organized, in part, by the Institute of Electrical and Electronics Engineers (IEEE), a U.S.-based engineering association.

Though it has not attracted as much global scrutiny as the March 2016 tournament between Alphabet's AlphaGo bot and a human Go champion, the recent Sejong competition is significant because the AI research community considers StarCraft a particularly difficult game for bots to master. Following AlphaGo's lopsided victory over Lee Sedol last year, and other AI achievements in chess and Atari video games, attention shifted to whether bots could also defeat humans in real-time games such as StarCraft... Executives at Alphabet's AI-focused division, DeepMind, have hinted that they are interested in organizing such a competition in the future.

The event wouldn't be much of a contest if it were held now. During the Sejong competition, Song, who ranks among the best StarCraft players globally, trounced all four bots involved in less than 27 minutes total. (The longest match lasted about 10 and a half minutes; the shortest, just four and a half.) That was true even though the bots were able to move much faster and control multiple tasks at the same time. At one point, the StarCraft bot developed in Norway was completing 19,000 actions per minute. Most professional StarCraft players can't make more than a few hundred moves a minute.

Comments

The tables have turned

[ShanghaiBill]

It used to be news when software beat humans at yet another game. Now it is news when we find a game that humans can still win.

Re:The tables have turned

[fbobraga]

Now it is news when we find a game that humans can still win.

and easily (this may explain why it is news...) - from the header:

all four bots involved in less than 27 minutes total. (The longest match lasted about 10 and a half minutes; the shortest, just four and a half.)

* former StarCraft player here: all games that lasted less than 10 minutes, that I can remember, all was extremely easy games...

Re:The tables have turned

[Maxo-Texas]

really insightful. wish i could mod you up.

now, if only jobs involved the total skill sets required by star craft, we could rest easy.

Re:The tables have turned

[Boronx]

Maybe not. These bots would probably beat 90+% of humans even if we were all trained in Starcraft.

Re: The tables have turned

[jandrese]

Seems strange that the bots were not able to deal with an early game rush, that seems like the area were bots could do quite well since the strategies are well known and it mostly comes down to playing out a script precisely.

Re: The tables have turned

[Zaelath]

That balance is exactly what a computer with basically perfect knowledge and zero reaction time should be better at than the human.

Re:The tables have turned

[NicknameUnavailable]

It used to be news when software beat humans at yet another game. Now it is news when we find a game that humans can still win.

How about a game I just made up, it's called "not being a filthy toaster." To win you have to not be a toaster. Someone write an article about this post now, I demand the front page.

Re: The tables have turned

[DevsVult]

Ah, but in this case the computer doesn't have perfect knowledge. You have to spend minerals on scouts to discover what the enemy is doing. The more scouts you have looking at the enemy, the less high-value units you have for attack and defence.

Re:The tables have turned

[gravewax]

I would settle for an AI that could pass even the basic turing test, we are yet to see any yet that can.

Re: The tables have turned

[Anonymous Coward]

There is a very easy explanation, the bots were badly coded.

Re:The tables have turned

[unknown_user_name]

Human players are as good as they are going to be. AI players are constantly improving. It took decades of progress for computers to win at chess. Many programmers and chess players assumed it couldn’t be done. Exponential improvements in hardware along with specialized algorithms and the relentless pace of machine learning overcame all obstacles. It will happen with Star Craft. The human players have peaked. The AI is just getting started

Re: The tables have turned

[Rakarra]

I remember the AI that came with Starcraft solved this problem quite nicely. The computer AI just always knew what buildings and what units you had. That made for practicing against a computer fairly hard because all the strategies that involved deception fell flat. Plus on the higher-difficulties, whenever you built a unit, the computer immediately knew, and put in production a unit to counter it. So the computer was often better at the macro game, but players were still better at the micro game.

Re: The tables have turned

[nasch]

They fixed that with Starcraft 2 and the AI doesn't know what you're doing any more.

Re: The tables have turned

[Zaelath]

I completely agree, but it does at least have perfect knowledge of it's own units etc. So you it can be hyper vigilant for sentry units killing it's scouts and still never miss a millisecond of build time as resources become available/deplete/etc.

I just suspect that if the bot makers were having their strategy written by the best human players they would be unassailable...

Re: The tables have turned

[jandrese]

Yeah, and it's very noticeable. In the original SC the AI was fairly predictable, you build and expansion base and the AI is going to mobilize the army to attack it just as your resource hub is finishing. In SC2 you can actually deny the computer information by intercepting its scouts and build expansion bases much more freely.

What's special about Starcraft?

[Opportunist]

I'm genuinely wondering, what makes Starcraft stand out? Is it something particular about this game or is it just the most well known RTS game, and AIs have generally a problem playing RTS games well?

Re:What's special about Starcraft?

[DontBeAMoran]

Probably the most popular RTS game right now. It's also pretty well balanced between playable races as I understand it.

Re: What's special about Starcraft?

[kodomo]

Mostly because it was a regular game at the wcg for more then a decade. Having professional players, leagues and channels devoted to it for longer than that. And about 10 years ago, it got hacked so a software could read the user variables and move the units, so the bot developing began before the machine defeated human on go, this is just the next natural step.

Re:What's special about Starcraft?

[Lanthanide]

Here's some general thoughts on what makes SC difficult for an AI.

Starcraft has areas of focus called 'macro' and 'micro'. Macro is base building, selecting which tech tree to advance down, upgrading, building your economy. Micro is controlling small groups or individual units and their position on the map and how they engage with enemy units. For a very rough idea, alternative terms would be macro = strategy, micro = tactics.

Someone who is excellent at macro can be destroyed by someone who is lousy at macro and excellent at micro, and vice versa. The top players are excellent at both macro and micro.

Scouting what your opponent is doing is a bit part of both your macro and micro strategies but in different ways. If you see the enemy has built unit factory X, then you better counter it by building Y (macro). If it's in map location Z, then you better make sure you get your units to position A to head them off (micro).

There are 3 races in SC so 6 possible match-ups include the 'mirror' matchups. If both players are random, then each player must scout their enemy to initially learn their race. Then there are known build-orders, so if you scout your enemy at time 2 minutes, and see X and Y, then you can conclude they are (probably) using strategy A. But if you get that exact same set of information at time 4 minutes, then concluding they are using strategy A may be a mistake.

Humans know these build orders (like expert players memorising important chess gambits), but the AI likely doesn't, so has to brute force everything. Brute-forcing is possible in chess, and thought impossible in Go. So Starcraft is an extension in this same area.

Re:What's special about Starcraft?

[Dutch Gun]

It's probably not just Starcraft. But since Starcraft is such a popular RTS, and is very mature as an esport, it makes a good metric for AI systems.

I'd bet the best computer bot would also get its ass kicked at Civilization by any competent player. And that's probably true of any other game with a complex enough ruleset which doesn't rely on sheer mechanical fitness (such as an FPS aimbot). Additionally, these are games with limited information about the state of the world. That is, your view is limited to your own units, and clouded by a fog-of-war system. In some sense, that's similar to poker, except that the world is too complex to take advantage of brute-force computational abilities in any way. So, you're left with the cleverness of your strategies, and computer AIs apparently aren't quite there yet.

Anyone who has ever played Civilization at higher than average difficulty settings knows the built-in AI cheats like crazy (or at least, as of a few versions ago - haven't played the newer versions). So, even the game developers apparently have a tough time programming an agent to play as competently as a human.

Re:What's special about Starcraft?

[Lanthanide]

big* part of both

Re:What's special about Starcraft?

[quantaman]

nothing special about starcraft. It is the style of game, it is much more complex to write smart AI for a RTS as the strategy is constantly evolving second by second. As bots improve they will easily demolish real players in RTS games as a human simply cannot compete with a machine that can make order of magnitude more game moves than you

I wonder how the bots do if both humans and bots are constrained to something like 60 moves a minute?

Re:What's special about Starcraft?

[Njorthbiatr]

I've played StarCraft competitively (note: not professionally) and follow the professional scene (yes there's still a pro scene). There's a special depth to Starcraft that other games lack.

Most of the ranks below top level are all about macro/micro mechanics. Macro (economy, building units, ensuring supply cap doesn't ever get hit) is easy for AI to do. Micro is more challenging but still a task that's better for computers than humans. Top level players are already so good at both of these that there's really only diminishing returns left for computers to gain an edge over them in. They're even so good as to know exactly if they'll win a fight between each unit; almost like they're subconsciously calculating the end game of any engagement automatically. You can even see them know exactly how many attacks it will take. It's freakishly superhuman.

They also have developed some sort of insane intuition. You can watch them play against each other and move units to locations or build defenses in places only seconds before they need them despite having zero knowledge of the impending threat (such as building turrets right when there's about to be a drop on your resource line). There's instances where the tactics become so deep that they manipulate "thinking ahead" to double or triple bluffs to create openings. To beat these players, you need to have a deep understanding of human motivations. Classic tricks such as hold position lurkers spell doom for computer opponents who need to understand where their enemy might be laying a trap for them, /if they have gone lurkers/. For all you know they're using stop lurkers to deceive you into thinking lurker/ling/hydra is their strategy, while they're in position to wipe out your workers with muta micro.

It's simply not a game where you can calculate the odds and win every game, because you lack sufficient information to calculate any meaningful conclusions.

Re:What's special about Starcraft?

[Kjella]

My guess is that they'd rather try something like StarCraft because the mechanics are more generic, like you collect resources, build units and attack enemies in quite free patterns whereas in Civilization the tactic is heavily driven by the game mechanics. Like, you do things in a particular order every time. I think Civilization would be easier for an AI to win with a balanced strategy and then win on micro-management.

Re:What's special about Starcraft?

[psycho12345]

Starcraft is fairly well balanced, despite being asymmetric. As such, there is no one guaranteed strategy to work each game, and not even against the same race. It is very old at this point, and mostly well understood. As a result, you can easily find players for the AI to fight, with a wide range of skill, ranging from noobs such as myself up to Korean Pros like Jaedong, who would dismantle most AI with his muta micro alone.

Re:What's special about Starcraft?

[west]

> I'm genuinely wondering, what makes Starcraft stand out?

Nothing. It's the other way around.

Traditional games like Chess and Go stand out by having massively fewer degrees of freedom than almost any video or even board game.

Not too diminish the advancements in AI, but games with relatively few choices each round are perfect for computers. In StarCraft (and in most video games), there are probably thousands of possible choices each frame.

Even if they got a Alpha Go like research budget, don't expect the Civilization 7 AI to start playing a game that's challenging to higher-level human play.

Re:What's special about Starcraft?

[Dutch Gun]

I think Civilization would be easier for an AI to win with a balanced strategy and then win on micro-management.

Were that the case, I don't think the game's built-in AI would have to cheat quite as much as it does. Granted, the in-game AI can only rely on the local machine's CPU power, but I still think you're underestimating how much strategic depth the game has. Micromanagement will only get you so far.

Re:What's special about Starcraft?

[gravewax]

it would make it more interesting but a computer would still ultimately have the advantage as even tiny speed advantages add up over time, an RTS like SC those tiny advantages stack up the longer the match runs. Some things computers are simply better at, a human cannot beat the response time of a computer unless the computer is artificially hobbled. It really comes down to how well the bots are programmed.

Re:What's special about Starcraft?

[javaman235]

Actually, you have to look at who's AIs are in the game, they are not all created equal. Facebook isn't playing at the level of Alphabet, and there's rumors of some companies out there at the edge of the defense world with some scary stuff the public isn't fully up to speed with yet. Have no doubt that AI can beat these players.

Re:What's special about Starcraft?

[nanoflower]

But the rules have changed. Not as in a complete overhaul but Blizzard regularly does rebalancing which involves changing the dmg and health of various units and even changing the rules governing the building of some units. That said Blizzard does document much of their changes so it's easy enough to add in that information into the AI's database.

Re: What's special about Starcraft?

[Lanthanide]

6 combinations, 9 permutations.

Home arena aside, Packers vs Raiders is the same game as Raiders vs Packers.

Re:What's special about Starcraft?

[Lanthanide]

But that's not "the best move", that's the point.

SC isn't a full-knowledge game. The AI has to spend resources in order to scout the enemy - resources that could be used on army production. So you have the extreme of an AI that spends all of it's money on scouts, and has no money left for an army. Or an AI that spends all of its money on army, and doesn't do any scouting.

In top-level competitive play, scouting is one of the make-or-break points in the game. You occasionally see a game where a player managers to kill the opponents first scout before that scout has been able to see enough of the enemy's technology. That player is now in a very very difficult spot - do they try and 'guess' what the opponent is doing, do they spend more money on another scout and get late information?

If the AI does scouting and sees building A, but the scout is killed before it can see building B, does it assume building B must exist, because that's the "most likely" build order? Lilke 70% of the time you see building A, you will also see building B, so just assume building B was there even though you didn't see it? If you make that hard assumption, 30% of the time you'll be wrong. A human may be able to pick up on contextual clues (the map, the opponents playing history on that map, their play style, what else they did in the tournament) to help make the right decision, but an AI player only has the data it saw with its probe to work off.

If the AI has a set of specific build orders that it specifically knows about, what does it do when it discovers a situation outside of those known build orders? If the human opponent knows that the AI is programmed only with specific build orders, then they can deliberately fake it out. A simple example is a "proxy barracks". Say you normally build a barracks (to produce basic troops) at time 3 minutes. AI scouts your base, doesn't see any sign of barracks or anything else - what does it do? Assume the player is not going to build any army at all? Turns out the player is actually building the barracks right next to the AI's base, but the AI incorrectly assumes there is no barracks at all, because they didn't see one. 90 seconds later, 3 marines turn up at the AI base, how does it react?

See also, this comment: https://games.slashdot.org/com...

Re:What's special about Starcraft?

[Altrag]

Go may only have a few dozen possibly moves per round, but the bot needs to be able to forecast multiple possibilities dozens or hundreds of rounds into the future to determine which move to make -- that's an exponential complexity and blows up really really fast. The challenge for Go (and Chess) bots is balancing the ability to forecast (power of the algorithm) against the amount of search tree pruning (speed of the algorithm.) Too much toward the former and the "player" becomes unresponsive. Too much toward the latter and the player becomes simply bad at the game. Really, this principle hasn't changed much since they first started building Chess bots. Hardware advances of course make the speed better, and now they're using neural nets to do the tree pruning because a) its more "intelligent" in the sense that it requires less direct human programming and b) it can potentially find ways to prune the tree that we humans would never dream of, thus making it at least in theory better than any possible direct-programmed bot, given enough training data.

StarCraft on the other hand is a very different style of game as you noted, but the problem isn't that there's so many possible things to do each game tick -- 99% of those things are irrelevant at any particular time and thus easily pruned out of the search tree in the first few layers. The big trick to StarCraft that the Google team was focusing on back when they announced this is the fact that you have imperfect knowledge -- that is, fog of war exists. Chess and Go and any other board game don't have that since they're based on games designed for humans sitting around a board. Things like card games have the hidden information aspect as well, but their search space is just too small, and outcomes too random (at least if its using a fair deck) to be an interesting project. Modern competitive like StarCraft, League of Legends, etc on the other hand have all of the hidden information going on but with very little randomness -- or at least, the randomness it does have is due to humans being unpredictable rather than being inherent to the game.

As for Civ7 or any other PVE game.. that's yet another whole ballpark. Competitive games work well because they're running on a level playing field meaning there's lots of relevant training data (recordings of real players) and they can also be trained by playing two copies of the AI against each other. Being the 'E' side of a PVE game has neither of those properties though, meaning that even if you create an AI it will be significantly harder to train it in any useful manner because there's almost no training data (you would have to iteratively set the bot parameters, find possibly hundreds or thousands of players willing to play against an absolutely terrible AI to generate some training, then plug that back in and iterate the whole thing again hundreds or thousands of times until the AI has learned enough to be useful.) And you can't just making a competing AI to be the player -- at least not up front -- or the two AI's will just run themselves into a loop that may have nothing to do with how a real player would react to anything. You might be able to get away with that after you've got enough real player data to use for training the player-emulating AI, but by that point you've probably already finished the majority of training your enemy AI anyway.

Re:What's special about Starcraft?

[Pulzar]

Even if they got a Alpha Go like research budget, don't expect the Civilization 7 AI to start playing a game that's challenging to higher-level human play.

That's a sad thought.

As the game gets more complex, the AI is only getting worse at handling it, making the latest installment the least captivating of them all :(.

And, yet, I still hope Civ 7 could be different. So, shut up. :)

Re:What's special about Starcraft?

[Altrag]

I was going to post something along those lines -- if we assume SC runs 20 game ticks per second, that's a maximum of 1200 APM even if Superman was playing.

But then I realized -- TFS wasn't strictly clear and its likely the 19000 APM games were probably coming from when it was two AIs training against each other rather than when it was playing against real opponents -- those games basically are run as fast as the hardware allows since they don't have to care about things like screen tearing, input device polling speed, allowing people with outdated hardware to still be able to compete, etc that the game has to be able to deal with when its a real person playing.

Re:What's special about Starcraft?

[im_thatoneguy]

what makes Starcraft stand out?

In a game of go a player has 19x19 (361) possible moves (spaces) per interval (turn).

Starcraft not only has a far larger board with the average 1v1 map being ~130x130 (16,900) possible moves (spaces) per interval (turn).

So just based on the board size the computer already has to calculate nearly 50x as many positions.

Now for those positions instead of 1 single move possible per space (place one token), you have multiple units.

Assuming we have 20 units and if we view Unit.Selected = [bool] as a variable for each unit then we have 2^20 (1,048,576) possible selections of those units. * multiple actions per unit (burrow, move, attack etc.) = let's say 3 million possible actions.

So for every AI frame\tick you've got 3m * 16,900 possible moves = ~ 45,000,000,000 moves in starcraft per turn vs 350 moves in Go. But it gets worse. A good starcraft player can make a move every let's say 2 seconds. For a game lasting 20 minutes that's 600 iterations\ticks\frames vs roughly 200 in Go.

Lastly the AI doesn't even know what the opponent is doing. Without a scouting unit the AI also has to infer everything that is happening based on a small fraction of the information it can know. So it has to strategize based on limited data unlike go where the whole board is visible.

So just random choices an AI could make is like 45B^650 compared to like 200^200 for Go. So while it's not realistic to compute Go like we computed chess, it's a whole 'nother level for Starcraft.

Re:What's special about Starcraft?

[nasch]

A good starcraft player can make a move every let's say 2 seconds.

Depending on what you mean by "good", the pros make 5 actions per second and better.

A few hundred moves a minute

[DontBeAMoran]

Most professional StarCraft players can't make more than a few hundred moves a minute.

Unless we're talking about just clicking non-stop to make your groups move a few pixels at a time, I'm pretty sure I can't manage more than a dozen moves or so per minute.

Re:A few hundred moves a minute

[damnbunni]

The game actually measures 'Effective Actions per Minute' and disregards spam clicks.

A 'casual player' is considered to be about 50 apm. 'proficient' players are about 150.

Selecting a building, selecting a unit, and training a unit would be three actions. Selecting a unit and commanding it to move would be two actions.

Re:A few hundred moves a minute

[kwoff]

The hands of God - Flash or 400 APM for example.

"Playability"

[fbobraga]

Starcraft have the best "playability" (or "fluid gameplay") in multi player RTSs in this time, and by far, besides all the technical limitations of the time: the game was very playable even with lowend graphics card / CPU and dialup connection

Re:"Playability"

[ma1wrbu5tr]

I first played this H2H against a buddy on a Pentium 166 Mhz with 64 MB RAM via a 33.6 kbps modem and STILL enjoy a round to this day. The original "craft" games were standard bearers for things to come and have yet to be overtaken.

not for long

[turkeydance]

enjoy your humanity fas long as it lasts

So, I'm putting a bet for Nov 27th, 2020

[Maxo-Texas]

At that point A.I. will completely own any non A.I. foe in star craft.

Because at that point properly configured A.I. will have been effectively playing for 200 years.

As with Go, it won't use existing strategies- humans will learn new strategies from playing against it.

And it won't cheat by playing 19,000 actions per turn.

Re:So, I'm putting a bet for Nov 27th, 2020

[Wootery]

humans will learn new strategies from playing against it

Unless those strategies aren't 'human-friendly', which they might well be.

The human mind is good at pattern-detection and heuristics, but is extremely bad at brute-force.

Re:So, I'm putting a bet for Nov 27th, 2020

[Altrag]

I just mentioned on a previous post, but I'm guessing the 19,000 APM is when two AIs are playing each other (and running as fast as the hardware allows) and its likely limited to a more realistic APM when playing against human opponents.

Re:So, I'm putting a bet for Nov 27th, 2020

[Maxo-Texas]

Alpha Go didn't use brute force.

Human go masters learned new plays that hadn't been discovered by any human player in 3000 years of playing Go.

Re:So, I'm putting a bet for Nov 27th, 2020

[Wootery]

Interesting - do you have a source? A quick Google didn't turn up anything promising.

Re:So, I'm putting a bet for Nov 27th, 2020

[Maxo-Texas]

alphago human masters new moves

Long list of results... here's a typical one.

https://www.wired.com/2016/03/...

SEOUL, SOUTH KOREA â" In Game Two, the Google machine made a move that no human ever would. And it was beautiful. As the world looked on, the move so perfectly demonstrated the enormously powerful and rather mysterious talents of modern artificial intelligence.

But in Game Four, the human made a move that no machine would ever expect. And it was beautiful too. Indeed, it was just as beautiful as the move from the Google machineâ"no less and no more. It showed that although machines are now capable of moments of genius, humans have hardly lost the ability to generate their own transcendent moments. And it seems that in the years to come, as we humans work with these machines, our genius will only grow in tandem with our creations.

Although machines are now capable of moments of genius, humans have hardly lost the ability to generate their own.

This week saw the end of the historic match between Lee Sedol, one of the world's best Go players, and AlphaGo, an artificially intelligent system designed by a team of researchers at DeepMind, a London AI lab now owned by Google. The machine claimed victory in the best-of-five series, winning four games and losing only one. It marked the first time a machine had beaten the very best at this ancient and enormously complex gameâ"a feat that, until recently, experts didn't expect would happen for another ten years.
===
alphago methodology

typical good result
https://www.dcine.com/2016/01/...

===
Interesting article on the new "alphago zero" which absolutely kicked alphago's ass.

https://medium.com/intuitionma...

Beat the previous version of AlphaGo (Final score: 100â"0).
Learn to perform this task from scratch, without learning from previous human knowledge (i.e. recorded game play).
World champion level Go playing in just 3 days of training.
Do so with an order of magnitude less neural networks ( 4 TPUs vs 48 TPUs).
Do this with less training data (3.9 million games vs 30 millions games). ...
Humans learn languages through metaphors and stories. The human strategies discovered in Go are referred to with names so as to be recognizable by a player. It could be possible that the human language of Go is inefficient in that it is unable to express more complex compound concepts. What AlphaGo Zero seems to be able to do is perform its moves in a way that satisfies multiple objectives at the same time. So humans and perhaps earlier versions of AlphaGo were constrained to a relatively linear way of thinking, while AlphaGo Zero was not encumbered with an inefficient language of strategy. It is also interesting that one may consider this a system that actually doesnâ(TM)t use the implicit bias that may reside in a language. David Silver, of DeepMind, has an even more bold claim:

Itâ(TM)s more powerful than previous approaches because by not using human data, or human expertise in any fashion, weâ(TM)ve removed the constraints of human knowledge and it is able to create knowledge itself.

===

You can also watch alpha go analysis on youtube. It's pretty dry but you typically get commentary by a go expert on the moves made.

---

The biggest limitation on A.I. is human. We don't have the right theories yet. We don't know how to formulate some problems for machine learning. In those areas where we do, machines rapidly exceed human capabilities.

btw, to me "brute

One thing constantly surprises me w/ these stories

[93 Escort Wagon]

For some reason I am always amazed to find out there are people who can make enough money to live on by playing video games.

And yes, I'm old.

That's only because Deepmind wasn't playing

[Sarusa]

Yes, human players can still beat second tier AIs from Facebook and universities.

But if you turned the AlphaGo Zero team on it it would dominate it in a couple months max. AGZ learned, from scratch, how to beat every human on the planet at Go every single time in 3 days.

Re:That's only because Deepmind wasn't playing

[Sarusa]

The same way meatbags 'solve' a starcraft game, except you have a million times more processing power?

AGZ can figure the best move at any time just from the current board setup. That should work in SC2 as well, though of course you need more TPUs and memory.

Re:That's only because Deepmind wasn't playing

[Sarusa]

Yeah, I wish I could upvote you on this.

AGZ doesn't seem to make any fantastic, earthshattering, amazing moves... it trades points, seemingly nothing special. Yet it inexorably ends up ahead on points to the relentless outcome where no human player can beat it.

It'd be the same with SC2. It wouldn't crush you at the start, but inevitably it would win.

Re:That's only because Deepmind wasn't playing

[guruevi]

Not really, Go and chess are full-knowledge games, you know at all points, where your opponent currently is, what his moves have been and where he can go. Not to say, there are only a very small numbers of paths you can take at any point in time.

StarCraft is a partial-knowledge game, you have to intuit where your opponent may be popping out what his intentions are based on very limited amounts of data and counter their strategy accordingly. There are also various ways of winning the game, you can starve your opponent, you can just go out and destroy him with superior force, annoy him continuously or simply execute a fast counterattack when their forces are away from the base in an offensive maneuver and most likely a combination of those things will win you the game. You can't just "guess" a solution because 90% of the times you will guess wrong, the game develops very differently based on tech trees your opponent chooses and choosing your own tech tree is a constant back and forth of trying to one-up your opponent.

This is really the worst situations for AI. There are no "common situations" as you have in chess or Go that you can just hard-code ideal responses to. AI's are still very poor at pattern recognition if the patterns aren't fully visible.

Re:That's only because Deepmind wasn't playing

[Sarusa]

This is such the God of the Gaps argument.

Oh you can't beat Checkers. Oh, you completely solved checkers? Well, you can't beat chess.

Uh, you beat chess, well you CERTAINLY can't beat Go. ... You beat Go? Well, you couldn't have beat Go without human training. ... ... ... You beat Go without human training, and it can beat every human on the planet every time?

Well, you can't beat Starcraft ha ha ha!

AIs have done pretty well in poker, which is another partial-knowledge game, and there's no reason to think a deep learning AI will do any worse at partial knowledge than a stupid meatbag.

Everything you just said about Starcraft 2 is what people said about Go till Go was handily trounced and they had to retreat to the next barricade.

Re:That's only because Deepmind wasn't playing

[gravewax]

Go is a full knowledge, limited move type game. You cannot compare learning the two as SC is micro management RTS with incomplete knowledge. maybe those guys would be able to make a great bot, but I would happily bet it would take a lot more than a couple of months and even longer before it dominated.

Re:That's only because Deepmind wasn't playing

[Anonymous Coward]

This is simply pure ignorance on your part, can't say I have EVER heard anyone with knowledge on the subject say making bots for any of those games that would beat any human was impossible. But checkers, chess and Go are all complete knowledge games, nothing hidden, it is not about whether you can beat SC or not. A bot can certainly beat any human should someone appropriately program it. hwoever you can't compare making bots for full knowledge limited move games against an incomplete knowledge, unlimited move game. The process for making the bot is completely dissimiliar and as such they certainly would not be coming up with a winning bot in just a couple of months, the ignorance here is your assertion that they could just translate the experience in writing it for Go to something like SC.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

Go and chess are full-knowledge games, you know at all points, where your opponent currently is, what his moves have been and where he can go. Not to say, there are only a very small numbers of paths you can take at any point in time.

For decades, computers have failed at Go because there are too many paths you can take. Since AlphaGo defeated Lee Sedol, people suddenly seem to think it's a relatively simple game.

Within a few years, we'll have an AI beat a human at StarCraft too. You'd better mount your goalposts on wheels.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

Go is a full knowledge, limited move type game.

Not really. You don't know what your opponent is planning.

Re:That's only because Deepmind wasn't playing

[goose-incarnated]

You don't seem to be aware that you're making a faith-based argument.

"They said we'd never achieve $FOO, and then we did. This proves we'd achieve $BAR" is a fundamentally flawed argument, regardless of what values you assign to FOO and BAR.

"They" said we'd never beat chess, "they" said we'd never beat Go, but "they" also said we'd never achieve time-travel into the past.

Oh, one more thing - no one said "we'd never beat Go": throughout the 90's I only ever heard "we'd never beat Go with current computers". They were right, too.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

"They said we'd never achieve $FOO, and then we did. This proves we'd achieve $BAR" is a fundamentally flawed argument, regardless of what values you assign to FOO and BAR.

Except in cases where FOO and BAR are essentially the same thing, but BAR is a bit further on the scale of size and complexity than FOO, and that we can reasonably expect our development of hardware and software to be able to tackle problems with greater scale and complexity in the future.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

How do you go about solving a Starcraft game where there are infinite possibility of moves?

At any point in time, the number of inputs is limited by the availability of controls. You can press a key, move the mouse, push a joystick or other controller in a finite number of ways.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

It's not actually infinite, just exponential.

Indeed. And Chess and Go are exponential too. The difference is the branching factor and depth of the tree you need to consider.

Re:That's only because Deepmind wasn't playing

[goose-incarnated]

"They said we'd never achieve $FOO, and then we did. This proves we'd achieve $BAR" is a fundamentally flawed argument, regardless of what values you assign to FOO and BAR.

Except in cases where FOO and BAR are essentially the same thing, but BAR is a bit further on the scale of size and complexity than FOO,

It's debatable whether "Win at Go" and "Win at Starcraft" are the same thing separated only by complexity, but let's be generous and assume that it is. We went from needing 30 x 120MHz CPUs to win at Chess (Deep Blue), to 1202 CPUs and 176 GPUs to win at Go (Alphago).

IOW, we used almost 1000x more resources to win at Go than at Chess.

For humans, at least, Go is roughly 2.5 times more complex than Chess . To address the 2.5 extra complexity going from Chess to Go, we used 1000x extra resources.

Starcraft, for humans at any rate, is a lot more than 2.5 times more complex than Go.

You're assuming that the required computational power/problem-solving scales log(x), like in this chart: as we get closer to the maximum computational power of the universe we'll solve more and more problems.

I'm more inclined to believe that the computational power/problem-solving chart looks like this (the right half only, obviously). IOW, the easy problems scale well, the hard problems are impossible.

Re:That's only because Deepmind wasn't playing

[SharpFang]

Except:

- you only know a part of the current board setup

- the more resources you dedicate to expanding your knowledge of board setup the less you have towards actual board domination.

- accurately predicting the board setup basing on what *should* be optimal currently is excessively difficult, as having a sub-optimal setup the opponent failed to predict is preferable to having optimal setup which the opponent predicted. That includes purposeful deceit.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

We went from needing 30 x 120MHz CPUs to win at Chess (Deep Blue), to 1202 CPUs and 176 GPUs to win at Go (Alphago).

AlphaGo Zero only uses 4 TPUs, and is much stronger than the 176 GPU version. It is also much stronger than the 48 TPU version that beat Lee Sedol, while only using a fraction of the space and power. If the goal is only to narrowly beat the human world champ, maybe 1 or 2 TPUs would suffice.

IOW, we used almost 1000x more resources to win at Go than at Chess.

AlphaGo Zero uses less power than Deep Blue did, and plays at a much higher level (comparing with best human players).

But the biggest problem with your analysis is that Chess is solved in a completely different way. Deep Blue didn't use neural nets, but relatively simple human-generated heuristics combined with deep brute force search. There have been some experiments with NN based chess programs, but so far, the results have not led to top-class performance. It is conceivable that a NN based chess program would require more resources than a Go program.

I agree that StarCraft would most likely require more processing units than Go, but at this point, that's all we can say.

How many more brain cells does a human StarCraft player use compared to a human Go player ? If it's a factor 12, then the AlphaGo Lee version, running on 48 TPUs may be able to do the job, given the proper algorithms.

Re:That's only because Deepmind wasn't playing

[goose-incarnated]

We went from needing 30 x 120MHz CPUs to win at Chess (Deep Blue), to 1202 CPUs and 176 GPUs to win at Go (Alphago).

AlphaGo Zero only uses 4 TPUs,

Which is much more powerful for the task than the 176 GPUs I stated above.

and is much stronger than the 176 GPU version. It is also much stronger than the 48 TPU version that beat Lee Sedol, while only using a fraction of the space and power.

It is much more computational power. There is a reason I kept saying "resources" and "computational power" instead of "electricity". You've sort of agreed that we've used multiple magnitudes of computing power to make the AI win.

If the goal is only to narrowly beat the human world champ, maybe 1 or 2 TPUs would suffice.

IOW, we used almost 1000x more resources to win at Go than at Chess.

AlphaGo Zero uses less power

I suggest you reread what I wrote in my original post. I never claimed that Alphago uses less electricity, but you imply that I made the electricity argument. I didn't.

I did claim "1000x more computational power".

than Deep Blue did, and plays at a much higher level (comparing with best human players).

But the biggest problem with your analysis is that Chess is solved in a completely different way. Deep Blue didn't use neural nets, but relatively simple human-generated heuristics combined with deep brute force search. There have been some experiments with NN based chess programs, but so far, the results have not led to top-class performance. It is conceivable that a NN based chess program would require more resources than a Go program.

I agree that StarCraft would most likely require more processing units than Go, but at this point, that's all we can say.

How many more brain cells does a human StarCraft player use compared to a human Go player ? If it's a factor 12, then the AlphaGo Lee version, running on 48 TPUs may be able to do the job, given the proper algorithms.

If we need 1000x more computational power for something a human finds 2.5x more complex, do you really think that we will only need 12x more computational power for something that humans find 12x more complex? My argument is that AI accomplishments are not scaling linearly, and they are fast hitting the wall with brute-force and NN.

Besides, as you say Deep Blue did not use NN, and the method Deep Blue used will likely not work well for AG anyway, resulting in the 2.5x complexity increase needing 1000x more computational power. What makes you think a 12x (or whatever) complexity is a tractable problem? All the evidence I've seen points to AI scaling being O(N^m) with N being the complexity.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

If we need 1000x more computational power for something a human finds 2.5x more complex, do you really think that we will only need 12x more computational power for something that humans find 12x more complex?

I was talking about the number of neurons in the human brain taking part in the game decision process. I think it's reasonable to assume that the scaling in the brain corresponds with the scaling in nodes in artificial neural nets. Perceptive complexity isn't a very good measure, I think. People find common tasks, like walking, simpler than playing StarCraft, but that could be because their brains are optimized for the first task, and not for the second. I think someone playing StarCraft uses a bigger part of their brain than someone playing Go, because there's more overlap between StarCraft and daily life, so more neurons can be recruited to join in the effort. But I highly doubt the difference is more than a factor 100.

Besides, as you say Deep Blue did not use NN, and the method Deep Blue used will likely not work well for AG anyway, resulting in the 2.5x complexity increase needing 1000x more computational power.

Because Deep Blue didn't use NN, I don't think it's useful to discuss the relative complexity increase.

What makes you think a 12x (or whatever) complexity is a tractable problem?

Because DeepMind already had a 12x bigger version running before.

All the evidence I've seen points to AI scaling being O(N^m) with N being the complexity.

I doubt it. Our neocortex is only twice as big as the chimpanzee's, and our total brain is only 3x the size, but we are capable of tasks that orders of magnitude more complex.

Re:That's only because Deepmind wasn't playing

[religionofpeas]

Apparently, you don't know what "limited move" means

Both games have limited moves. You can control StarCraft with a mouse and keyboard, and each of those input devices only has a fixed range of inputs. The difference is only in scale. In StarCraft, you send a few inputs per second, and in Go you can take a minute.

Re:That's only because Deepmind wasn't playing

[goose-incarnated]

If we need 1000x more computational power for something a human finds 2.5x more complex, do you really think that we will only need 12x more computational power for something that humans find 12x more complex?

I was talking about the number of neurons in the human brain taking part in the game decision process. I think it's reasonable to assume that the scaling in the brain corresponds with the scaling in nodes in artificial neural nets. Perceptive complexity isn't a very good measure, I think. People find common tasks, like walking, simpler than playing StarCraft, but that could be because their brains are optimized for the first task, and not for the second. I think someone playing StarCraft uses a bigger part of their brain than someone playing Go, because there's more overlap between StarCraft and daily life, so more neurons can be recruited to join in the effort. But I highly doubt the difference is more than a factor 100.

Besides, as you say Deep Blue did not use NN, and the method Deep Blue used will likely not work well for AG anyway, resulting in the 2.5x complexity increase needing 1000x more computational power.

Because Deep Blue didn't use NN, I don't think it's useful to discuss the relative complexity increase.

What makes you think a 12x (or whatever) complexity is a tractable problem?

Because DeepMind already had a 12x bigger version running before.

All the evidence I've seen points to AI scaling being O(N^m) with N being the complexity.

I doubt it. Our neocortex is only twice as big as the chimpanzee's, and our total brain is only 3x the size, but we are capable of tasks that orders of magnitude more complex.

Yeah, but human and chimp brains aren't AI, and don't work the same way that NNs do. NNs scale linearly, for example, but biological brains do not. NN are not digital representations of biological brains. If they were by now we'd have machines with the sentience sentience at least a cockroach, but we do not.

Re:That's only because Deepmind wasn't playing

[guruevi]

Go, even though the number of moves for an entire game are very high, only has a limited set of moves that wouldn't outright lose you the game. The players also have an overview of the state of the board at all times.

Eventually, there will be "AI" that can beat people at StarCraft, but that doesn't mean it will be any time soon. So far, its been trying with brute force, simply trying to be better than human at controlling units and trying to abuse certain properties of the game.

Re:That's only because Deepmind wasn't playing

[chispito]

StarCraft is a partial-knowledge game, you have to intuit where your opponent may be popping out what his intentions are based on very limited amounts of data and counter their strategy accordingly.

Churn a bunch of replays through the AI and I bet it could learn these strategies and tactics pretty quickly.

Re:That's only because Deepmind wasn't playing

[gravewax]

no, but you know EXACTLY what moves they have made, you do not have that luxury in SC

Re:That's only because Deepmind wasn't playing

[guruevi]

We've had AI SC games for more than a decade, they have improved yet not quite as stellar.

Woot!

[ma1wrbu5tr]

Go team "Smelly bags of mostly water!" Way to humiliate those dry silicates!

Re:Woot!

[Chris Mattern]

That's "ugly bags of mostly water", thank you very much!

Re:One thing constantly surprises me w/ these stor

[93 Escort Wagon]

I’d take offense at your comment, but it’s time for my nap. So maybe later, if I remember.

Re:Mass shooting in Texas

[Joce640k]

I don't get this at all.

It's Texas. Wasn't everybody in the building carrying some sort of concealed firearm to defend themselves with, including the preacher?

Re:One thing constantly surprises me w/ these stor

[Altrag]

People make enough money to live on by throwing or kicking balls around, slapping rubber disks toward a net, punching each other in the face, etc.

People making money by "playing" games isn't exactly new. Of course just like physical sports, when you're playing at the professional level, its no longer just a game -- its a job and requires huge amounts of time and effort to maintain your skills at the peak level. Possibly even more effort than physical sports since in addition to maintaining their "fitness," players also have to continually keep up with changing game mechanics and metas as all major esports games are continually getting new units/avatars, balance tweaks, etc.