20 ms·
Facebook Quietly Enters StarCraft War for AI Bots
- scj 9y agoHere's a link if you want to see CherryPi vs. Purplewave (two bots mentioned). https://youtube.com/watch?v=rTf_aL0hrgo https://youtube.com/watch?v=rTf_aL0hrgo
- 9935c101ab17a66 9y agoThanks for the link, really cool to see the AIs at work.
- Strilanc 9y agoThat'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.
- pdimitar 9y agoSee my root-level comment in this thread: https://news.ycombinator.com/item?id=15437177 https://news.ycombinator.com/item?id=15437177
- dgant 9y agoHi 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.
- dgant 9y agoHi! Author of PurpleWave here. A bit late to the party but happy to answer questions.
- vk23 9y agoHow far away are the AIs from consistently beating humans? What are the biggest problems you are facing right now?
- dgant 9y agoThe best AIs right now would beat above average folks from the public. But as you start approaching anyone who resembles a professional, bots rapidly turn into practice dummies. The biggest problems I'm personally facing at the moment are midgame macro with imperfect information -- how many Gateways do I need to be producing from before I can safely take a third base or transition to endgame tech? -- and engaging Terran mech armies. It's very hard to know when or how to engage a Terran mech army, and you usually only get one chance to get it right. Dithering results in bleeding units you can't afford to lose. The biggest problem in general is that rules-based approaches are asymptoticly approaching the limit of their capabilities. Either of the problems I've described are big undertakings but only scratch the surface of what PurpleWave needs to advance. Machine learning is the way forward. But so far real time strategy games have proved remarkably impenetrable. Consider that professional players often place their initial buildings to provide Scarab-proof escape routes for their workers against Reaver drops that arrive ten minutes later. Such placements are only effective if you know to react to Reaver drops by funneling your workers through them to glitch the Scarabs. How are you going to let your bot learn that on its own? There's a lot of work to do.
- vm 9y agoMisleading 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 9y agoAs 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.
- AstralStorm 9y agoThe 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...
- comicjk 9y agoWhat that tells me is that they haven't yet come up with the right featurization - that is, the function that maps input data into the actual neural network node values. The appropriate featurization would include the time information but reduce its dimensionality by hard-coding some basic assumptions, of the kind that humans presumably make when processing the same data.
- yingxie3 9y agoI think these guys (and most people using deep models) try to avoid hand-crafted features as much as possible.
- infecto 9y agoAnyone have good resources on how to get into this? Would love to learn more about self-learning in regards to starcraft.
- make3 9y agothe easier way : watch the course vids here : Berkeley CS294 Deep Reinforcement Learning Fall 2017: http://www.youtube.com/playlist?list=PLkFD6_40KJIznC9CDbVTjAF2oyt8_VAe3 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 https://arxiv.org/abs/1701.07274 progressively implementing subparts in Python and openai gym https://gym.openai.com/read-only.html https://gym.openai.com/read-only.html
- etiam 9y agostarcraftai.com ?
- oneweekwonder 9y agoblizzard have a github account with a couple of py libs for sc2 and other games. https://github.com/blizzard https://github.com/blizzard
- pdimitar 9y agoIRC FreeNode, channel #BWAPI. Most of the authors of the bots competing 24/7 in https://twitch.tv/sscait 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://sscaitournament.com/ - https://github.com/dgant/purplewave https://github.com/dgant/purplewave
- make3 9y agoomg what a stupid title from wired. wired showing again that they are a sensationalist piece of internet garbage
- Jyaif 9y agoYeah, I'm disappointed their articles keep popping up on HN.
- tree_of_item 9y agoThey came in sixth. That's what happened. Exactly what is so sensationalist here?
- m777z 9y ago6/28 is a very respectable finish, though I was surprised that the top 3 finishers were just individuals, not even teams.
- make3 9y agothey'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 9y agoWhere can I watch other Starcraft Bots play each other? Anyone have links to the matches?
- shagie 9y agoTwitch is where I would be digging around. https://twitch.tv/sscait https://twitch.tv/sscait is one such with commentary.
- arcticfox 9y agoIf anyone is interested in similar RTS AI play but for Age of Empires 2: https://youtu.be/ePgEUSazsos https://youtu.be/ePgEUSazsos
- alexmorenodev 9y agoThere is StarCraft Artificial Intelligence Tournament on youtube. https://www.youtube.com/user/certicky/ https://www.youtube.com/user/certicky/
- horsecaptin 9y agoIs it fair to flag all articles that bait / mislead readers?
- rising-sky 9y agoWeak article but still surprised there's no mention of Open AI's foray into Dota 2: https://blog.openai.com/dota-2/ https://blog.openai.com/dota-2/
- crsv 9y agoI'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 9y 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.
- gdb 9y ago(I work at OpenAI.) Correct, we've been playing a number of pros using the same bot played at TI. We do have a stronger version which is just two days more of training (gets a 70% win rate vs the one at TI), but haven't seen a need to test it out. We'll likely do a blog post in upcoming weeks with more stats and commentary from the pros; would be curious what people would like to know! Incidentally, the various exploits that people used are all similar to how we actually develop the bot. We try to find areas of the strategy space it hasn't explored, and then make a small tweak to encourage it to explore that area. Lots of progress comes from removing hardcoded restrictions, which are nice to get started. So the fact there exist exploits wasn't surprising to us — what would be surprising would be exploits we couldn't fix. 1v1 has always been a proof-of-concept for us. The fun part is really 5v5, which is what we're working on now (and hiring for! ping me if you're interested: gdb@openai.com).
- huac 9y 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 9y agoThey need a research paper to show they're not pulling numbers from their ass.
- yeahbutbut 9y agoDo people really click on ads on purpose?
- stagger87 9y agoYes...?
- thephyber 9y agoPerhaps 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 9y agoDoes 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.
- ncds42 9y agoUnless there are in-depth APM limitations on the AI, there's no contest for AI vs. A human.
- stale2002 9y agoUmm, 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.
- thephyber 9y agoWhat 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.
- ncds42 9y agoYour 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.
- dgant 9y agoBrood War competitions have been going on for seven years, with individual bots getting years of development. If beating professionals was easy it'd have been done by now. It turns out that being really good at narrow micromanagement situations doesn't add up to winning complete games. StarCraft is messy and difficult.
- lfowles 9y agoOnly 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/facebook-careers/dinoisland https://github.com/robertdimarco/puzzles/tree/master/faceboo...
- amenghra 9y agoFun 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/Student-hacker-penetrated-Facebook.html http://www.telegraph.co.uk/technology/facebook/8708392/Stude...
- Fiahil 9y agoAh! 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 https://github.com/fiahil/Zappy
- yters 9y agoIf AI weren't a religion, people would realize it is a joke by now.
- CyberDildonics 9y agoQuietly means paying for an article to be written in Wired now?
- Merem 9y 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 http://sscaitournament.com/index.php?action=tutorial An older post on this topic: http://www.teamliquid.net/blogs/485544-intro-to-scbw-ai-development 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 http://wiki.teamliquid.net/starcraft/SSCAIT
- bfung 9y agoHmm, 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.
- XR0CSWV3h3kZWg 9y agoHonestly 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.
- haeffin 9y agoSC2 ticks way faster than SC1 though, which you don't notice much as a human, but you might change your strategy when designing the ai components.
- XR0CSWV3h3kZWg 9y agoAlso I'd imagine the amount of time you spend making sure that units will path efficiently is way less!
- pdimitar 9y agoOne of the bots authors participating in SSCAIT (https://twitch.tv/sscait https://twitch.tv/sscait) once sent me this link: http://wiki.teamliquid.net/starcraft/Micro_Training_Maps 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- 9y agoDave 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."
- dwighttk 9y ago"Facebook Quietly..." headline on Wired... um. nothing after that could be true.