4 ms·

Generalized reasoning in strategy games AI is an especially difficult problem for machine learning. IRL, top StarCraft players routinely model their opponents m

by wamatt 10y ago

Generalized reasoning in strategy games AI is an especially difficult problem for machine learning. IRL, top StarCraft players routinely model their opponents mental states and psychology to create an edge. So perhaps it's worth pointing out, that this paper specifically addresses a sub-problem of Starcraft play, micromanagement ('micro') [1] The game engine runs at 24 frames per second. (As an aside, 'frames' in this context likely does not map to physical FPS of the display). >We ran all the following experiments with a skip_frames of 9 (meaning that we take about 2.6 actions per unit per second). The research team found that attempting to move at a superhuman pace (eg one action every frame), resulted in a subpar performance and hyper-parameterization indicated 2.6 to be an ideal action per second. In context, this translates to an APM of 156. Or, roughly half that of professional Korean e-athletes. [2] [1] https://en.wikipedia.org/wiki/Micromanagement_(gameplay) https://en.wikipedia.org/wiki/Micromanagement_(gameplay) [2] https://en.wikipedia.org/wiki/Actions_per_minute https://en.wikipedia.org/wiki/Actions_per_minute

aab0 10y ago
"The researchers found that attempting to move at a superhuman pace (eg one action every frame), resulted in a subpar performance." Moving at extremely fine-grained timesteps can make learning much more difficult, because now a reward arrives millions of timesteps delayed rather than hundreds or thousands. It's like trying to teach a NN to compose piano music by starting down at the 1ms raw audio level. This is part of why audio synthesis was so difficult up until recently with DeepMind's WaveNet. In theory, being able to move every frame should enable extremely superhuman performance, but in practice, you can't learn your way there. So often people will chunk data to make it easier to learn the higher-level concepts: operate on words, rather than characters, for example.
- raus22 10y ago
  Why not go the other way and decrease the actions per minute so you learn the overall point of the game , And with each game the actions per minute increases.
  - daveguy 10y ago
    That sounds like an interesting research angle. The thing about AI research is there are so many open ends there are essentially unlimited research options. If you can pose it as a problem and identify a reasonable programming approach then you have an avenue for AI research. Deep Learning isn't the end of AI research. It is the beginning.
  - comex 10y ago
    Or maybe extend the traditional categories of macro and micro with another one, call it 'nano'... the micro agent indicates where each unit ought to be in 9 frames, and the nano agent figures out how to take them there. Since the timescale is so short, the agent could brute-force enumerate possible moves to some extent and figure out which is optimal, like chess AI. Or use a separate network. I guess that's inelegant when a deep network already has its own concept of fine-grained versus coarse-grained layers, and should be able to do this on its own with the right training method.
jcl 10y ago
APM numbers are actions-per-minute, not the actions-per-unit-per-minute measurement in which the AI's performance is reported. It's not clear how the two map to each other -- the AI may well be exceeding human performance when multiple unit groups are considered.
- wamatt 10y ago
  That's a good point and not really correct to conflate the two. It would then seem to suggest, that forces with more than say, 5 units (780 average sustained APM) and above, would likely be getting into super human territory.
  - rfrey 10y ago
    I am not a gamer... but I have to ask. 780 APM is possible with human players? 13 actions per second? I can't come close to clicking my mouse that fast. I don't think I can blink that fast.
    - Merem 10y ago
      Yes, it's possible. Bear in mind that not just clicks count though. July managed a peak APM of 818 in Stracraft: Brood War in an official game. The average APM of a match is usually between 300 and 500, also depending on the race that is played.
      - empath75 10y ago
        A lot of that is 'spamming', where they keep their hands moving and doing things even if they're not particularly meaningful, just so they can keep the pace of the eye/brain/hands feedback loop going.
        BirdieNZ 10y ago
        Maybe early on it is spamming, but later on in the game most actions because useful actions. Having said that, some APM measurements only tracked effective APM (EAPM), which was clicks that actually achieved some result rather than merely selecting a unit.

by wamatt 10y ago

aab0 10y ago

"The researchers found that attempting to move at a superhuman pace (eg one action every frame), resulted in a subpar performance." Moving at extremely fine-grained timesteps can make learning much more difficult, because now a reward arrives millions of timesteps delayed rather than hundreds or thousands. It's like trying to teach a NN to compose piano music by starting down at the 1ms raw audio level. This is part of why audio synthesis was so difficult up until recently with DeepMind's WaveNet. In theory, being able to move every frame should enable extremely superhuman performance, but in practice, you can't learn your way there. So often people will chunk data to make it easier to learn the higher-level concepts: operate on words, rather than characters, for example.

raus22 10y ago
Why not go the other way and decrease the actions per minute so you learn the overall point of the game , And with each game the actions per minute increases.
- daveguy 10y ago
  That sounds like an interesting research angle. The thing about AI research is there are so many open ends there are essentially unlimited research options. If you can pose it as a problem and identify a reasonable programming approach then you have an avenue for AI research. Deep Learning isn't the end of AI research. It is the beginning.
- comex 10y ago
  Or maybe extend the traditional categories of macro and micro with another one, call it 'nano'... the micro agent indicates where each unit ought to be in 9 frames, and the nano agent figures out how to take them there. Since the timescale is so short, the agent could brute-force enumerate possible moves to some extent and figure out which is optimal, like chess AI. Or use a separate network. I guess that's inelegant when a deep network already has its own concept of fine-grained versus coarse-grained layers, and should be able to do this on its own with the right training method.

jcl 10y ago

APM numbers are actions-per-minute, not the actions-per-unit-per-minute measurement in which the AI's performance is reported. It's not clear how the two map to each other -- the AI may well be exceeding human performance when multiple unit groups are considered.

wamatt 10y ago
That's a good point and not really correct to conflate the two. It would then seem to suggest, that forces with more than say, 5 units (780 average sustained APM) and above, would likely be getting into super human territory.
- rfrey 10y ago
  I am not a gamer... but I have to ask. 780 APM is possible with human players? 13 actions per second? I can't come close to clicking my mouse that fast. I don't think I can blink that fast.
  - Merem 10y ago
    Yes, it's possible. Bear in mind that not just clicks count though. July managed a peak APM of 818 in Stracraft: Brood War in an official game. The average APM of a match is usually between 300 and 500, also depending on the race that is played.
    - empath75 10y ago
      A lot of that is 'spamming', where they keep their hands moving and doing things even if they're not particularly meaningful, just so they can keep the pace of the eye/brain/hands feedback loop going.
      - BirdieNZ 10y ago
        Maybe early on it is spamming, but later on in the game most actions because useful actions. Having said that, some APM measurements only tracked effective APM (EAPM), which was clicks that actually achieved some result rather than merely selecting a unit.