Facebook Quietly Enters StarCraft War for AI Bots

https://www.wired.com/story/facebook-quietly-enters-starcraft-war-for-ai-bots-and-loses/

by sus_007 • 9 years ago

230 121 9 years ago

wired.com

scj2• 9 years ago

Here's a link if you want to see CherryPi vs. Purplewave (two bots mentioned).

https://youtube.com/watch?v=rTf_aL0hrgo

9935c101ab17a66• 9 years ago

Thanks for the link, really cool to see the AIs at work.

Strilanc• 9 years ago

That's surprisingly frustrating to watch. Protoss throwing away its army of zealots with terrible positioning, zerg wasting time attacking buildings instead of sealing the deal, probes walking by enemy forces that ignore them as they mine minerals...

Which is not to say that there weren't impressive parts. I don't think a human could stutter step that many dragoons that well. But there's clearly a lot left to do in terms of tactics.

pdimitar4• 9 years ago

See my root-level comment in this thread: https://news.ycombinator.com/item?id=15437177

dgant6• 9 years ago

Hi Strilanc. Author of PurpleWave here.

That particular scene at the sunken colonies was equally frustrating to watch. I fixed the most likely cause of that behavior shortly after the tournament deadline.

But for sure, you're correct. There's a lot left to do on all fronts. There's noticeable progress every year -- none of last year's entries finished much over 50% -- and the growing research community should help accelerate that even more.

dgant6• 9 years ago

Hi! Author of PurpleWave here. A bit late to the party but happy to answer questions.

vk23• 9 years ago

How far away are the AIs from consistently beating humans? What are the biggest problems you are facing right now?

scj2• 9 years ago

Both Iron Bot and krais0 have one helluva Terran mech (tank/Goliath/Vulture) push. Do you think a Protoss AI can defeat it?

vm2• 9 years ago

Misleading title. FB seems to be taking a self-learning approach, whereas hobbyists are defining bot rules themselves. Hobbyists are winning at the moment with an approach that doesn't scale. Self-learning has a better chance of dominating the field.

Edit: HN mods updated title to be less click-baity (thanks). Earlier it said "FACEBOOK QUIETLY ENTERS STARCRAFT WAR FOR AI BOTS, AND LOSES". The "and loses" at the end was misleading.

toxik• 9 years ago

As someone who has implemented quite a few reinforcement learning techniques and seen their limitations, I would be surprised if RL could overcome handcrafting for SC any day soon.

AstralStorm2• 9 years ago

The main way the AI bots have problems is with timing. The neutral networks used have no way of encoding time dependent actions in a reasonable way. (As opposed to say a fuzzy decision tree with explicit time input.) And if you try to explicitly include it, curse of dimensionality strikes back hard.

Both absolute and relative timing have to be handled. And relative since specific salient action...

Plus the real reward is very sparse. Say, crippling mineral production early may or may not snowball. Likewise being a unit or two up...

unpwn• 9 years ago

Why do you say that? The dota2 bot open ai did earlier this year seemed pretty convincing and similar...

smrtinsert• 9 years ago

Factually correct title.

tree_of_item3• 9 years ago

Misleading? How? It's the title of the actual article, and it's what happened. Just because they'll probably win in the future because their approach is more sound, doesn't mean they didn't lose now.

vm2• 9 years ago

Makes it seem like they lost in an AI bot battle. But they are in fact creating AI bots that compete against rule-based bots. Big difference.

infecto• 9 years ago

Anyone have good resources on how to get into this? Would love to learn more about self-learning in regards to starcraft.

make33• 9 years ago

the easier way : watch the course vids here : Berkeley CS294 Deep Reinforcement Learning Fall 2017: http://www.youtube.com/playlist?list=PLkFD6_40KJIznC9CDbVTjA...

harder more complete way: read this "Deep Reinforcement Learning: An Overview", https://arxiv.org/abs/1701.07274 progressively implementing subparts in Python and openai gym https://gym.openai.com/read-only.html

etiam• 9 years ago

starcraftai.com ?

oneweekwonder• 9 years ago

blizzard have a github account with a couple of py libs for sc2 and other games.

https://github.com/blizzard

pdimitar4• 9 years ago

IRC FreeNode, channel #BWAPI. Most of the authors of the bots competing 24/7 in https://twitch.tv/sscait hang around in various times of the day and night. They can provide you with plenty of entry material.

Few links:

- https://sscaitournament.com/

- https://github.com/dgant/purplewave

make33• 9 years ago

omg what a stupid title from wired. wired showing again that they are a sensationalist piece of internet garbage

Jyaif• 9 years ago

Yeah, I'm disappointed their articles keep popping up on HN.

dymk• 9 years ago

If the title has an anti corporate slant, especially FB/Google, it will continue to be upvoted on HN regardless of actual content

tree_of_item3• 9 years ago

They came in sixth. That's what happened. Exactly what is so sensationalist here?

m777z• 9 years ago

6/28 is a very respectable finish, though I was surprised that the top 3 finishers were just individuals, not even teams.

make33• 9 years ago

they're not learned ai, they're built in classical hard coded logic perfected through the years, with lots of domain specific knowledge from the users. that's a much easier way to produce something that does okay than to innovate with a new ai scheme, but importantly it also has no chance of ever beating a player who's at the professional level of play, let alone a top one

cgb223• 9 years ago

Where can I watch other Starcraft Bots play each other?

Anyone have links to the matches?

shagie• 9 years ago

Twitch is where I would be digging around. https://twitch.tv/sscait is one such with commentary.

arcticfox• 9 years ago

If anyone is interested in similar RTS AI play but for Age of Empires 2:

https://youtu.be/ePgEUSazsos

alexmorenodev• 9 years ago

There is StarCraft Artificial Intelligence Tournament on youtube. https://www.youtube.com/user/certicky/

horsecaptin• 9 years ago

Is it fair to flag all articles that bait / mislead readers?

rising-sky2• 9 years ago

Weak article but still surprised there's no mention of Open AI's foray into Dota 2: https://blog.openai.com/dota-2/

crsv• 9 years ago

I'm actually glad no one mentioned this because this event in and of itself was overhyped publicity. The Open AI bot had hard coded behaviors and was defeated handily within hours of being available to the public. It's since been iterated upon and there's still a handful of pros who are regularly beating it WITHOUT exploiting the holes in the logic. This is effectively just as weak as the article regarding facebook.

hsrada• 9 years ago

and was defeated handily within hours of being available to the public.

The public exploited tricks to beat it. They did not beat it 'handily'.

Afterwards, the pros who do beat it only manage to do so ~2-3 times for every 100 games played. I believe they have been playing the same version that was shown at The International and not an iterated version.

rising-sky2• 9 years ago

Interesting... I was not aware of this, thanks

huac• 9 years ago

A Microsoft research paper on machine learning this year said that improving predictions of when a user will click on an ad by just 0.1 percent would yield hundreds of millions of dollars in new revenue.

Do you really need a 'research paper on machine learning' to understand that?

icebraining• 9 years ago

They need a research paper to show they're not pulling numbers from their ass.

yeahbutbut• 9 years ago

Do people really click on ads on purpose?

stagger87• 9 years ago

Yes...?

thephyber4• 9 years ago

Perhaps that's the trick -- if you make an ad look less like an ad, it's 0.1% more likely to convert.

And Google still hasn't figured out how to best their search advertising network business, so I assume that people click on ads, even if I'm not in that demographic.

You should really watch a child play a game that has interstitial ads. It's quite obvious that they often click on ads because they want to learn more (maybe not fully convert, but intentionally click).

CamelCaseName• 9 years ago

Does anyone have the link to this paper?

I am curious if that millions in revenue is for Microsoft (not surprising) or advertisers (more interesting) - I would love to read through their thought process either way.

ncds422• 9 years ago

Unless there are in-depth APM limitations on the AI, there's no contest for AI vs. A human.

stale20022• 9 years ago

Umm, they have human vs AI challenges in starcraft and the AIs are vastly outclassed by humans. No APM limits needed.

AIs have a long way to go to beat good human starcraft players.

thephyber4• 9 years ago

What you say is true right now, but in the long term, the gp is correct unless we hit an AI asymptote which prevents AI from approaching human StarCraft capabilities.

Just from a pure economy standpoint, any computer process has quite an advantage from just optimizing action queues and keeping idle workers working.

ncds422• 9 years ago

Your last comment is, frankly, complete bullshit.

People have already created TAS programs that can take out infinite numbers of enemies with minimal stock (ex: medvac/tank vs. infinite ultralisks). Or have zerglings that perfect 1 siege blast per 1 zergling splitting.

As far as micro, AI already has proven capability to absolutely dominate human micro, like not even close.

As far as build paths and macro decisions, AI isn't there yet but all it takes is one player and one programmer to come up with an in-the-middle and well-rounded build path that doesn't lose to any cheese; Sacrifice some economy to just have an army at all times_ and the AI will micro dominate the rest in extremely, humanly impossible army trades (I mean winning a 40 stock vs. 200 stock army battle).

Honestly just imagine having ONE mutalisk perfectly micro all-game, never receive death damage that just outputs as much damage as humanly possible at every angle. And you could have all 20-140 of your army stock doing this at all points in the game.

No contest. Just hasn't had time devoted to it yet.

lfowles• 9 years ago

Only slightly related, but does anyone remember the code puzzles that FB had as a filter for the interview process? Specifically one (jurassic island?) where you would connect to a shared server and take control of a dinosaur on the island. Eventually when you gathered enough energy to lay an egg, you could spawn another connection to control that dino.

I had no hope of having a competitive score on the leaderboard, but I loved playing around with that little world.

Edit: Found more info on dinosaur island: https://github.com/robertdimarco/puzzles/tree/master/faceboo...

amenghra• 9 years ago

Fun story, there was a pretty serious vulnerability in there [1].

Lessons learned: don't run CTFs "too close" to your real infrastructure.

[1] http://www.telegraph.co.uk/technology/facebook/8708392/Stude...

Fiahil• 9 years ago

Ah! I created a very similar game as a school project (a few years ago) ! It was a 2nd year project, pardon the messy commits : https://github.com/fiahil/Zappy

yters• 9 years ago

If AI weren't a religion, people would realize it is a joke by now.

CyberDildonics• 9 years ago

Quietly means paying for an article to be written in Wired now?

Merem• 9 years ago

"The contest Facebook entered, like most AI research in the area, used an older version of StarCraft, which is considered equally difficult for software to master."

I'm surprised this one hasn't been called out yet but Starcraft: Brood War and Starcraft II are completely different games, not just different versions.

And apart from what was already mentioned regarding information on the scene, this link is pretty useful: http://sscaitournament.com/index.php?action=tutorial

An older post on this topic: http://www.teamliquid.net/blogs/485544-intro-to-scbw-ai-deve...

Facebook: facebook.com/groups/bwapi

As for SSCAIT in particular: http://wiki.teamliquid.net/starcraft/SSCAIT

bfung3• 9 years ago

Hmm, are they truly completely different games?

From a high level, it seems like the elements for an AI to learn are mostly the same: Resource gathering and optimization, map/location battle optimizations, building/technology decisions...

It's not like Starcraft: verison X is different than Starcraft: version Y is different the way Starcraft and Hearthstone are different games.

XR0CSWV3h3kZWg2• 9 years ago

Honestly from an AI perspective SC:BW and SC2 are probably more similar than from a human's perspective. 12+ unit selection, multi building selection & intelligent worker rallies aren't huge improvements for AI.

lazzlazzlazz• 9 years ago

They are truly completely different games.

FlorianRappl• 9 years ago

I agree with the sentiment that for a human the difference is much larger than for a machine, however, taking into consideration that a huge amount of the AI in SCBW needs to be on the micro (i.e., in-battle) execution which is - for the most part - handled by the in-game AI in SC2 I feel both games are different enough on AI development. Some basics are definitely shared, but beyond a certain level the two games go in different directions.

One thing to add is the openBW project (http://www.openbw.com). It is truly amazing: The guys there managed to reverse engineer the full SCBW engine (I thought that's nearly impossible, as the whole flow depends a lot on implementation details, some of the details are even considered bugs, which turned out to be just the right implementation to balance the game perfectly...).

taw55• 9 years ago

I haven’t played either, but I know for a fact that they use different pathfinding methods (grid vs navmesh, II has swarm behavior).

david-cako• 9 years ago

AI has a significant edge in micro and mechanics in BW that is much less significant in SC2.

dsjoerg• 9 years ago

I'd say about as similar as football and soccer. Both involve getting the ball thingy into the opponent's end-of-field thingy. Indeed lots of common skills.

pdimitar4• 9 years ago

One of the bots authors participating in SSCAIT (https://twitch.tv/sscait) once sent me this link:

http://wiki.teamliquid.net/starcraft/Micro_Training_Maps

They are supposed to be mini maps where your bot can train separate aspects of the game on a much smaller scale.

Extremely useful if you're an aspiring bot author.

2Pacalypse-• 9 years ago

Dave Churchill, who is heavily involved in the StarCraft AI scene, posted this in the StarCraft AI Facebook group:

"Hey everyone, I just wanted to make a quick statement about this article that was posted in WIRED. If Tom had told me that his intention for this article was to focus on the loss of ANY bot, Facebook or not, I would not have given him the information that I did to write this article. I am normally a huge fan of Tom's AI articles, but I think that this is a shameful clickbait title that focuses on the wrong side of the story.

It turns out that there may actually be a bug with how Facebook's serialization library interacted with our tournament environment, causing their strategy selection to not work properly, causing a huge drop in performance for their bot. We will be investigating this over the next few days to see if it was my mistake in setting things up that caused this, or if maybe something as simple as a windows firewall setting to be at fault. Unfortunately with competitions like this, competitors do not have access to our tournament environment / machines to fully 100% test all of their code on, and sometimes issues like this can arise. I appreciate Facebook's effort in this competition and applaud them for actually having the balls to submit an open source bot, knowing the scrutiny they would face if it wasn't already superhuman. It's a real shame that this is what media chooses to focus on. Tom also asked me what time the results would officially be made public so that the article could be released at the same time, but instead the article ended up being published 7am the next morning so everyone could wake up to the headline :( Also the word "quietly" implies some sort of sneaking around on the part of Facebook, when their registration has been publicly visible for months.

Note: There is nothing factually incorrect in this article, I just believe its focus is in poor taste, and hurts the SCAI community as a whole, and unfairly singles out a single competitor."

late2part• 9 years ago

Fail early, fail often. The article was lame. I encourage the FB team to keep going. No one cares if you fail - we just care if you win. Go Gadget Go!

randartie• 9 years ago

Link to facebook group? A search for 'Starcraft AI' doesn't bring up any results on facebook for me.

badosu• 9 years ago

Not sure if it's the same one I followed for some time, but maybe it's this one: https://sscaitournament.com/

dgant6• 9 years ago

https://www.facebook.com/groups/bwapi

dwighttk• 9 years ago

"Facebook Quietly..." headline on Wired... um. nothing after that could be true.

Facebook Quietly Enters StarCraft War for AI Bots

Related Stories