Quote:
Originally Posted by grizy
The thing with programs like Deepmind is as soon as it understands the rules, it only takes a few iterations (human intervention iterations) before they become superhuman.
What you wrote describes AlphaZero, but I'm not sure if they are using that or a completely different tool/strategy to tackle StarCraft. Having imperfect information (not to mention an exponentially larger number of variables) probably has a big impact on how much of their previous work in chess/go applies to SC2.
Quote:
Originally Posted by grizy
My bet is Terran. Mostly because perfect micro of tanks and marines seem to be the most superhuman thing with most potential to increase win equity. I mean... how do you defend against someone that can perfectly micro 3, 4, even 5 drops simultaneously?
Also worth noting - I think they (DeepMind) want their bots to have to perform under human-like constraints, i.e. a few hundred APM and not 20k. While superhuman multitasking (drops/drop defense) is still within those parameters, I think things like tank micromanagement or superhuman marine splits are not. Even if they were, it would kinda be cheating - the goal is to
outsmart humans, not just take advantage of a faster clock speed.
Quote:
Originally Posted by PokerPlayingGamble
in b4 deep mind picks protoss and cannon rushes
I remarked to a coworker that it would be absolutely hilarious if, after dozens of iterations of training and playing against itself to distill the absolute perfect StarCraft bot, it comes out on stage and shows the world the most unexploitable strategy of all: cannons.