deepmind's alphastar beats pros in starcraft ii

[WinPooh]

But realtime factor is always in favor of computer. I am almost sure that decent 2300-2400 chess engine can easily beat Carlsen in bullet. But it can't beat him at longer time control.

[Nay Lin Tun]

But realtime factor is always in favor of computer. I am almost sure that decent 2300-2400 chess engine can easily beat Carlsen in bullet. But it can't beat him at longer time control.

They also restrict AI activity that is practically impossible for human. For example , human cant do two commands on two areas of screen simultaneously (battling in North pole of screen and doing economy in home base). They add about 15-20 ms lag/punishment for AI in changing point of view of screen/triggering area.

[duncan]

https://deepmind.com/blog/alphastar-mas ... rcraft-ii/

Have not seen all of it. Are there any plans to improve it and beat the world champion?

[Werewolf]

Great presentation. Wonder why they trained on Protoss? Shorter matches?

They actually explain it in the demonstration video.

It must have been the bit when I went to the toilet then because I watched all of it. What was the reason?

[henk2]

Results in chess and go were much more impressive to me. Maybe because I always knew that computer can "click his mouse" faster than any human.

They restrict AI's action per minute for fair play with human. There is details discussion in reddit.
Afaik, human did 450 APM(action per minute) and AI did 250 APM in average. But people said only 50% of human APM would probably be effective action, cos usually human repeatedly click mouse or shift key several times without effective action or double or triple click for the same action.

The effective APM (EPM) of a human is at most 33% of their real APM.

Alphastar was also allowed to "store" actions in a way, as it micro'd a three way battle at 1500 apm.
They averaged it to 250 apm, but in short spikes Alphastar was allowed to use much more.

[Uri Blass]

But realtime factor is always in favor of computer. I am almost sure that decent 2300-2400 chess engine can easily beat Carlsen in bullet. But it can't beat him at longer time control.

I am not sure that you are right.

It may be interesting to try stockfish that is slowed down by a significant factor so it is only 2300-2400 at tournament time control against humans to see what it the rating of it in bullet against humans.

[WinPooh]

But realtime factor is always in favor of computer. I am almost sure that decent 2300-2400 chess engine can easily beat Carlsen in bullet. But it can't beat him at longer time control.

I am not sure that you are right.

It may be interesting to try stockfish that is slowed down by a significant factor so it is only 2300-2400 at tournament time control against humans to see what it the rating of it in bullet against humans.

Or we can take Carlsen's or Nakamura's bullet games against human opponents and check them for blunders with Phalanx or Fritz 5.

Remember, Kasparov lost to Chess Genius in rapid (not bullet) in 1994, and almost lost blitz tournament to Fritz 3 next or previous year. Both programs were much weaker than 2700 in classical chess.

[Uri Blass]

But realtime factor is always in favor of computer. I am almost sure that decent 2300-2400 chess engine can easily beat Carlsen in bullet. But it can't beat him at longer time control.

I am not sure that you are right.

It may be interesting to try stockfish that is slowed down by a significant factor so it is only 2300-2400 at tournament time control against humans to see what it the rating of it in bullet against humans.

Or we can take Carlsen's or Nakamura's bullet games against human opponents and check them for blunders with Phalanx or Fritz 5.

Remember, Kasparov lost to Chess Genius in rapid (not bullet) in 1994, and almost lost blitz tournament to Fritz 3 next or previous year. Both programs were much weaker than 2700 in classical chess.

It is known that old software like Fritz5 or Genius3 were stronger at faster time control relative to humans.
I am not sure if it continue to be correct with new software that part of the advantage of it is being able to earn more from time (from the same rating point).

[Joost Buijs]

Interesting analysis:

https://medium.com/@aleksipietikinen/an ... 02fb8344d6

[Daniel Shawul]

Also there is no lookahead search (mcts) for the starcraft ai since the branching factor is much much higher than Go.
It just plays what it thinks is best strategy without opponent modeling.