4 ms·

An Analysis on Deepmind’s Starcraft 2 AI’s Superhuman Speed

16 points by sytelus 8y ago

MiddleEndian 8y ago
To me, having a starcraft AI win via super high APM is like having a robot beat Usain Bolt in a sprint, but the robot has wheels as its base instead of legs. I'm more interested when they can do it with the same physical constraints, like if that humanoid-shaped Boston Dynamics robot (famous for being harassed after a box was knocked out of its hands) could outsprint Usain Bolt.
- karmakaze 8y ago
  An updated version of AlphaStar that didn't cheat in the same way with the camera as the previous versions did lost to MaNa. So I expect to see more games with additional restrictions to come closer to meaningful losses, and perhaps victories.
OrwellianChild 8y ago
Does anyone more familiar with SC2 analysis know how they reduce APM to EPM? If there is a systematic way to remove spam-clicks in training data, it seems like it would be relatively straightforward to train AlphaStar on only the "meaningful" clicks, thereby eliminating this issue...
- kalid 8y ago
  That's a great point. I think you could train a network on what clicks were meaningful or not by looking at whether the future game state was changed by the presence of the click. (I.e., if you click to attack 4 times, but clicks 2 3 and 4 don't change the impact of the battle when they are removed, they were likely spam.)
greendestiny_re 8y ago
I find the linked article incomplete and the way DeepMind framed the story disingenuous. There was no mention of the fact the DeepMind team produced 5 versions of AlphaStar and pitted them all against both TLO and MaNa without telling them they're playing against essentially five players. This turned what human players thought was best-of-5 into 5 rounds of best-of-1, which is not how any SC2 tournaments are played. My guess is that this was done on purpose to mask the fact each agent was highly specialized in one particular strategy, which the humans could adapt to and counter it.

16 points by sytelus 8y ago

To me, having a starcraft AI win via super high APM is like having a robot beat Usain Bolt in a sprint, but the robot has wheels as its base instead of legs. I'm more interested when they can do it with the same physical constraints, like if that humanoid-shaped Boston Dynamics robot (famous for being harassed after a box was knocked out of its hands) could outsprint Usain Bolt.

karmakaze 8y ago
An updated version of AlphaStar that didn't cheat in the same way with the camera as the previous versions did lost to MaNa. So I expect to see more games with additional restrictions to come closer to meaningful losses, and perhaps victories.

OrwellianChild 8y ago

Does anyone more familiar with SC2 analysis know how they reduce APM to EPM? If there is a systematic way to remove spam-clicks in training data, it seems like it would be relatively straightforward to train AlphaStar on only the "meaningful" clicks, thereby eliminating this issue...

kalid 8y ago
That's a great point. I think you could train a network on what clicks were meaningful or not by looking at whether the future game state was changed by the presence of the click. (I.e., if you click to attack 4 times, but clicks 2 3 and 4 don't change the impact of the battle when they are removed, they were likely spam.)

greendestiny_re 8y ago

I find the linked article incomplete and the way DeepMind framed the story disingenuous. There was no mention of the fact the DeepMind team produced 5 versions of AlphaStar and pitted them all against both TLO and MaNa without telling them they're playing against essentially five players. This turned what human players thought was best-of-5 into 5 rounds of best-of-1, which is not how any SC2 tournaments are played. My guess is that this was done on purpose to mask the fact each agent was highly specialized in one particular strategy, which the humans could adapt to and counter it.