|
Paper : arxiv.org
Now before everyone gets super excited :
'From a machine learning point of view, StarCraft provides an ideal environment to study the control of multiple agents at large scale, and also an opportunity to define tasks of increasing difficulty, from micromanagement, which concerns the short-term, low-level control of fighting units during battles, to long-term strategic and hierarchical planning under uncertainty. While building a controller for the full game based on machine learning is out-of-reach for current methods, we propose, as a first step, to study reinforcement learning algorithms in micromanagement scenarios in StarCraft. '
Give them five years.
|
This isn't DeepMind, it's Facebook.
|
Shouldn't this in the SC2 section?
|
Yeah, for some reason I always think Soumith Chintala works at Deepmind ( admins could you change the title ?). It's a carbon copy of methods those guys used anyway. I need to sleep.
They are looking at BW specifically. The article mentions dragoons and wraiths on page 4.
|
United Kingdom13775 Posts
Good for them I guess. Not the first group interested in using Starcraft for AI research, and won't be the last.
I expect them to ultimately succeed at performing well-optimized micro movements under very narrow scenarios of gameplay.
|
Croatia9500 Posts
Changed the title.
And indeed, they're using BW (and BWAPI) which is pretty exciting.
Here is a short (non-technical) summary that might be more interesting to people here:
- In all tasks, we control all units from one side, and the opponent (built-in AI in the experiments) is attacking us
- Tasks are: 5 marines vs 5 marines, 15 marines vs 16, 3 zealots + 2 dragoons on both sides, 15 wraiths vs 17, other scenarios with X marines vs Y marines and X wraiths vs Y wraiths
- In 15 marines vs 16 scenario, the advantage is on our player’s side - 81% win rate
- whereas in 15 wraiths vs 17 is hard - 10% win rate
- 5 marines vs 5 marines scanario - 85% win rate
- 3 zealots + 2 dragoons vs 3 zealots + 2 dragoons - 48% win rate
- The subsequent experiments will include self-play (training and evaluation), multi-map training (training more generic models), and more complex scenarios which include several types of advanced units with actions other than move and attack
- Finally, the goal of playing full games of StarCraft should not get lost, so future scenarios would also include the actions of “recruiting” units (deciding which types of unit to use), as well as make use of them
So looks like they're just getting started! Hopefully they come far enough to start participating in the BW AI tournaments.
|
Very interesting, and strange... seems like a weird publicity move to choose broodwar given how little attention it gets from the mainstream nowadays, but still super cool. maybe it'll bring more people to bw when they realize that facebook is tinkering with it
|
United Kingdom13775 Posts
On September 13 2016 16:58 Endymion wrote:
Very interesting, and strange... seems like a weird publicity move to choose broodwar given how little attention it gets from the mainstream nowadays, but still super cool. maybe it'll bring more people to bw when they realize that facebook is tinkering with it
Technically simpler (game engine), with more previous work to piggyback on. It's a logical choice given that BW has been theoretically interesting to the AI community for a while now.
|
i know there's always being a lot of work done on broodwar AI in the community but i didn't realize that it was large enough to actually get recognized by a company like facebook for investment, everyone in the botting community must be pretty excited about the level of these new eyes on them
|
United Kingdom13775 Posts
Just another group looking to look into just another interesting challenge in AI. Ultimately that's all it is.
|
They have in their midst one of the 'founding fathers' of broodwar AI, Gabriel Synnaeve, I for one will be following this project closely 
|
United Kingdom13775 Posts
|
2Pacalypse, thanks for this. I'll post a detailed summary of the article when I get a sec, just wanted to give everyone a heads-up that this is happening.
|
next facebook game will be a starcraft 1x1 against their AI.
|
United Kingdom13775 Posts
On September 13 2016 16:58 Endymion wrote:
Very interesting, and strange... seems like a weird publicity move to choose broodwar given how little attention it gets from the mainstream nowadays, but still super cool. maybe it'll bring more people to bw when they realize that facebook is tinkering with it
So I have a better answer for you as to why, if you read that U Alberta link. Basically the reason is BWAPI and how Blizzard does not allow any similar software to modify SC2, per the EULA. So it's a legal issue.
|
On September 14 2016 06:28 LegalLord wrote:Show nested quote +On September 13 2016 16:58 Endymion wrote:
Very interesting, and strange... seems like a weird publicity move to choose broodwar given how little attention it gets from the mainstream nowadays, but still super cool. maybe it'll bring more people to bw when they realize that facebook is tinkering with it So I have a better answer for you as to why, if you read that U Alberta link. Basically the reason is BWAPI and how Blizzard does not allow any similar software to modify SC2, per the EULA. So it's a legal issue.
This for sure, as well as in general, when training neural networks on pixels, you want as little of those as possible ( as it makes the statistical estimation process easier ), both in terms of resolution and FPS. So if the same game was available to play on an old school Nintendo versus a Playstation 4 HD, you'd want to use the former for AI training purposes. BW makes more sense by that logic too.
|
Hmm.... kind of crazy and interesting
|
bw's eula allows 3rd party input but sc2's doesn't?? i would have expected both of them to not allow modification.. i wonder if hdbw's eula will be the same as sc2s
|
On September 14 2016 17:10 Endymion wrote:
bw's eula allows 3rd party input but sc2's doesn't?? i would have expected both of them to not allow modification.. i wonder if hdbw's eula will be the same as sc2s
Well, back when BW eula was written, DLL injection wasn't widespread knowledge. So since DLL injection doesn't alter the .exe itself it is technically allowed in that sense.
But even if Blizzard tried to stop it they can't. Since BW has private servers, and you can play via DirectIP or LAN with Hamachi. Whereas in SC2 you have to use Battle.net to play multiplayer.
|
United Kingdom13775 Posts
So I got a good chance to read through the paper. They're basically applying more modern machine learning techniques to very simple micro scenarios as described by 2Pac. They talk about how Starcraft represents a rather interesting challenge for AI because of its large, almost world-like state space with uncertain information, but that they're basically going to abstract all of that away by focusing on the tiny problem of reliably fighting the built-in AI under simplified scenarios (map is visible etc).
On the one hand, it's nice that they're not arrogant enough to believe that they're going to solve a game that is well beyond the ability of modern AI capability to play well (computers have no sense of strategic thinking and you can't just search a tree like in chess/Go to find an optimal/high-quality move because Starcraft just isn't so statically defined). And their learning algorithm approach is certainly interesting. I'd like to see what they come up with, on video. On the other hand, even by AI standards beating a built-in AI (i.e. no micro whatsoever) under simplified circumstances is nothing special. I look forward to seeing what they come up with, but I think it's more relevant to their own research work than to Starcraft AI-making. I might be wrong but I'm going to take the "skeptic by default" position here until they prove otherwise; hype usually leads nowhere unless there's evidence that it's not just fluff.
|
|
|
|
|
|
EPT
