Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepMind and Blizzard Open StarCraft II as an AI Research Environment

deepmind.com

Post reply on HN
Print view View original

61–70 of 283 posts

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#61

Post by gwern »

wfunction wrote: post #44
gwern wrote: post #40

Earlier quoted context omitted.

Until today's release of the headless Linux client, you still had to run the full StarCraft program, which gets expensive fast. And it massively complicates the workflow to have to play through every game serially to recreate the state rather than simply reading random rows of data from a 300GB dataframe on disk.

Oh I see, thanks, I didn't know. But man, 300 GB per game sounds completely nuts!

No, total. For comparison they quote the replay files at what was it, 5GB? It's a classic space-time tradeoff, but in deep learning right now, hard drives are far cheaper than CPUs/GPUs. Playing out the games as you need individual datapoints would probably be at least twice as slow, while anyone can easily store 300GB these days.

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#63

Post by sooper »

wfunction wrote: post #53
jzymbaluk wrote: post #49

Earlier quoted context omitted.

I believe the 400GB is the total amount for the 65000 different game replays

Oh but how does that work? That's ~6 MB per game which sounds like just a list of actions rather than precomputed data per frame. Is it compressed somehow?

"The full dataset after compression is 365 GB, 1535 million frames, and 496 million player actions." - Yes

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#64

Post by hossbeast »

CuriouslyC wrote: post #52
captainmuon wrote: post #43

Earlier quoted context omitted.

> SC2 got rid of a lot of that You think so? My impression is that SC2 had a lot more of repetitive tasks you had to do. E.g. wall off the ramp, send a worker scouting, ... and you have to perform certain actions every X seconds (like using chrono boost). A lot or mastering the game is rote learning, and polishing a build order. Another big part is constantly scouting and reacting to what the enemy is doing. Due to t…

I really like the idea of a helper AI that takes high level commands and handles the manual dexterity aspect of the game. That would make RTS games a lot more attractive for someone like me who prefers simple, thoughtful games rather than complex games with a steep learning curve.

Civilization?

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#65

Post by bytefactory »

gwern wrote: post #55
bytefactory wrote: post #46

Earlier quoted context omitted.

Ooh, this'll be interesting to see, as with AlphaGo, a lot of people disputed the "experts believed it would take ~10 more years" claim retrospectively. With SC2, no AI even comes close to beating even a silver level player, so even a 5 year timeline seems really soon. Let's see if DeepMind can beat it! What's your totally unscientific guess, Gwern?

I think it is doable in under 5 years, but this critically depends on the resources invested by DM and other DL orgs. Deep RL is hugely demanding of computational resources to iterate your designs - for example, the first AlphaGo took something like 3 GPU-years to train it once (2 or 3 months parallelized); however, with much more iteration, DM was able to get Master's from-scratch training down to under 1 month. Now…

> (It might not even be as complex as people think ...

Yeah, I suspect you're right. Eliezer was alluding to this with the AlphaGo victory as well:

> ... Human neural intelligence is not that complicated and current algorithms are touching on keystone, foundational aspects of it. https://www.facebook.com/yudkowsky/posts/10153914357214228?p...

I can't decide if I would be bummed or excited if that turns out to be the case. On the one hand, we'd be that much closer to AGI. On the other, we'd be continuing down the path of brute-forcing intelligence, rather than depending on those elegant, serendipitous breakthroughs that much of human progress has been built on.

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#66

Post by cjbprime »

hitekker wrote: post #59

This seems all in good fun but I wonder if it's come too late. Starcraft 2 is at its twilight. The biggest leagues of South Korea have disbanded. [1] The prolific progamers who transitioned to Starcraft 2 have gone back to Broodwar. [2] Blizzard itself has scrubbed all references to Starcraft 2 on the very home page of Starcraft. [3] Except for the twitter embed, it has only only one "2" character... in the copyright…

I don't quite agree, FWIW.

SC2 does seem to be at its twilight in Korea, and I agree progamers and fans there are super interested in Remastered.

But I don't think Remastered will be very popular outside KR. The SC2 "war chest" promo appears to have made more money than expected, as measured by hitting its funding ceiling within a few days.

So I don't think it's "Remastered replaces SC2", I think it's a divergence into KR playing Remastered and non-KR playing SC2, and the number of progamers and players doesn't have to be zero-sum: it could enlarge the population playing either game, too.

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#67

Post by cjbprime »

siegecraft wrote: post #4

The API Blizzard is exposing is really nice. Sadly most of the advantages AI had in SC1 were just due to the fact that an automated process could micro-manage the tasks the game didn't automate for you (a lot of boring, repetitive work). SC2 got rid of a lot of that while still allowing room for innovative and overpowered tactics to be discovered (MarineKing's insane marine micro, SlayerS killing everyone with blue f…

ByuN's reaper micro!

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#68

Post by artursapek »

aerovistae wrote: post #41
ipnon wrote: post #2

Any predictions for how long it will take for an AI to win against the world's best player?

Awhile. This just isn't like Go or Chess. The gap from perfect information to imperfect information is quite a chasm, and from turn-based to real-time is even more vast. I play Age of Empires 2 semi-competitively, and I just can't imagine the research progress that would have to be made for a pro to lose to an APM-limited AI agent. So much of the game comes down to intuiting what your opponent is planning without bei…

Speaking of AoE 2, I really hope this research ends up benefiting that game. The AI in it has always been so bad but it's still my favorite game of all time.

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#69

Post by cjbprime »

aerovistae wrote: post #41
ipnon wrote: post #2

Any predictions for how long it will take for an AI to win against the world's best player?

Awhile. This just isn't like Go or Chess. The gap from perfect information to imperfect information is quite a chasm, and from turn-based to real-time is even more vast. I play Age of Empires 2 semi-competitively, and I just can't imagine the research progress that would have to be made for a pro to lose to an APM-limited AI agent. So much of the game comes down to intuiting what your opponent is planning without bei…

On the other hand, bots are starting to beat professionals at (thousands of hands of repeated) Poker, so I think we can't say that imperfect information is something that's especially intractable for maching learning algorithms.

http://spectrum.ieee.org/automaton/robotics/artificial-intel...

Re: DeepMind and Blizzard Open StarCraft II as an AI Research Environment

#70

Post by antisthenes »

captainmuon wrote: post #43
siegecraft wrote: post #4

The API Blizzard is exposing is really nice. Sadly most of the advantages AI had in SC1 were just due to the fact that an automated process could micro-manage the tasks the game didn't automate for you (a lot of boring, repetitive work). SC2 got rid of a lot of that while still allowing room for innovative and overpowered tactics to be discovered (MarineKing's insane marine micro, SlayerS killing everyone with blue f…

> SC2 got rid of a lot of that You think so? My impression is that SC2 had a lot more of repetitive tasks you had to do. E.g. wall off the ramp, send a worker scouting, ... and you have to perform certain actions every X seconds (like using chrono boost). A lot or mastering the game is rote learning, and polishing a build order. Another big part is constantly scouting and reacting to what the enemy is doing. Due to t…

> You think so?

It's an absolute fact. Here are a few things that reduced the necessary repetitive action count considerably compared to SC1:

1. Larger control groups 2. Worker queues (including sending them to the resource patches) 3. Smart casting (no longer have to select individual units to correctly chain cast certain AOE spells) 4. Pathfinding actually works now so no need to click 15 times to get a unit where you want it to go.

> I played SC1 more on LAN, and there wasn't all the metagame going on.

Oh, I see. So you never played competitively?

Post reply on HN