TorchCraft – StarCraft: Brood War AI(facebook.ai)
facebook.ai
TorchCraft – StarCraft: Brood War AI
https://facebook.ai/developers/tools/torchcraft
17 comments
The other side of the coin is that the game engine for StarCraft II is signifcantly smoother, and so its dynamics are easier for models to learn. Brood war has all kinds of weird quirks and dynamics, and some "bugs" are part of the game system, so it's much harder for computers to reason about some of these kinds of things.
Brood War definitely requires a much higher APM, but the quirks of the game could make it much harder to decipher what is the correct action. It'd be interesting to see if/when the AI would discover things like muta clumping and if it would uncover new micro techniques...that would be very cool.
Does anyone know if this is headless or if it needs to render the game elements? Can I inject macros with this? That is, can I be playing a game, cue the AI to do some micro for me then let me keep playing? I assume there are other tools to do this but I would think they would interact with my mouse cursor rather than the BWAPI.
There are a few ways to run Brood War headlessly. And if not headlessly, sped up as fast as your CPU can process frames (and Brood War is very fast -- it ran on Pentium 90s!). TorchCraft is compatible with any Brood War client that supports BWAPI, including the headless ones.
BWAPI supports optional user interaction: https://bwapi.github.io/class_b_w_a_p_i_1_1_game.html#a4db25... and https://bwapi.github.io/class_b_w_a_p_i_1_1_game.html#a4db25... -- someone actually just used BWAPI to make an automated real-time coach: http://www.teamliquid.net/forum/brood-war/533143-coach-ai-ha...
BWAPI supports optional user interaction: https://bwapi.github.io/class_b_w_a_p_i_1_1_game.html#a4db25... and https://bwapi.github.io/class_b_w_a_p_i_1_1_game.html#a4db25... -- someone actually just used BWAPI to make an automated real-time coach: http://www.teamliquid.net/forum/brood-war/533143-coach-ai-ha...
Am I going to start getting targeted ads based on my build order and actions per minute?
This type of cheap humor doesn't really add much to the conversation. I presume you don't actually believe that to be true, and are just trying to bring up Facebook's entirely unrelated privacy issues (which have already been discussed on HN ad nauseam).
How about discussing interesting AI tooling and research instead?
How about discussing interesting AI tooling and research instead?
Ehh, what I found funny as commentary on how far out of left field this is for Facebook. They want AI people to help them make money from advertising, so in effect they are doing this as advertising which actually does put it into their preview.
From the right angle you can see it as one of those in soviet Russia jokes.
From the right angle you can see it as one of those in soviet Russia jokes.
This type of comment doesn't really add much to the conversation. How about instead of presuming what someone means you respond to them as if they're being serious.
We can definitely talk about the implications of AI like this on the current funding model of the internet. Or we could just all be annoying to each other.
We can definitely talk about the implications of AI like this on the current funding model of the internet. Or we could just all be annoying to each other.
I thought it was pretty funny... it's ok to lighten up a bit.
Being funny is really not what people come to the hacker news comments for. While I appreciate a good chuckle as much as the next person, without more nuanced insight, or thought provoking perspective that I didn't see myself, or sharing relevant technical expertise, it's definitely wasting my time, and WILL get a downvote from me. I would be reddit for the snarky comments and to "lighten up a bit" on the same topic.
why so serious?
Yea we come to the hacker news comments for the bikeshedding and west coast political bias!
Why not ? It could suggest people to watch on twitch who are pros or there is a startup at YC that targets people who play competitively .
Gamers are a lucrative market. We have GPGPUs thanks to us. I trained a deep learning network because my brother bought video cards. DJB even states to optimize your algorithms with gamers in mind .
Gamers are a lucrative market. We have GPGPUs thanks to us. I trained a deep learning network because my brother bought video cards. DJB even states to optimize your algorithms with gamers in mind .
Made me think: if an AI analyzed my choice of race, my playstyle, build order execution, APM, etc. and suggested a particular skilled pro I'm most likely to learn from, I'd be really happy. This is the holy grail of targeted ads - very timely, very relevant. Alas, the past decade of experience with targeted ads doesn't give me much hope.
Are there any resulting data we can look at? Did anybody find wrote about actually running this for a while and seeing anything cool emerge?
+ As others have mentioned: Brood War is significantly cheaper to run (can be headless, runs well on ancient CPU), making it more amenable to the self-play with massive numbers of games approach
+ Brood War benefits significantly (arguably more than SC2) from skilled micromanagement, which is arguably easier to exploit as it doesn't require long-term/high level planning (but requires a generous actions per minute cap)
+ This API doesn't require the AI to also do simple computer vision/recognition of the game (vs. DeepMind's simplified graphical representation)
+ This API interfaces with game actions directly (BWAPI) vs. mouse/keyboard level actions
It'll be interesting to see which problem (Brood War vs. StarCraft II) gets solved to a compelling degree with learning first.