Algorithms Discover Build Order From Hell

This is mad. Over on StarCraft 2 forum Teamliquid, a poster who goes by Lomilar has been talking about a program he’s coded called EvolutionChamber. It uses genetic algorithms to find powerful build orders, meaning his program takes a population of build orders, kills off the useless ones, and has the most successful ones reproduce asexually to create a new population, which tests itself again, and so on. I’m taking all this from this blog post by programmer Louis Brandy, wherein he breaks down what Lomilar’s done so that lay folk can understand it.

EvolutionChamber’s already come up with one ludicrous build order, which I’ve posted beneath the jump.

10 extractor-trick to 11
11 overlord
11 spawning pool
15 extractor
16 queen (stop drones here)
18 overlord
18 roach warren
17 overlord (yes, two)
spawn-larva on queen when she pops
roach x7

Which will get you a 7 roach rush in less time than anybody thought possible. “Extractor-trick” refers to telling a drone to build an extractor, building a drone, then cancelling the extractor’s construction to get the drone back, making your population of drones 11 out of a maximum of 10.

To summarise Louis' blog post, this is interesting because:

(1) This is such a devastating build order that for a while the discussion on Battle.net centred around whether Protoss could survive it even if they knew it was coming. Turns out they can, but only with very specific counter-strategies or if the person they're playing fails to fulfil the potential of the build order.

(2) Despite its effectiveness, this build order wasn’t known to the SC2 community. It’s quickly become labelled as a “cheese build,” though, meaning... well, it works.

(3) The extractor trick used by the program has been tested by real players, and was thought to be “economically inferior” compared to just buying an overlord and increasing your population cap. However, using this trick the program created a build order that made 7 roaches in less time than the standard roach rush strategy produces 5.

(4) ...and yet while the strategy neglects to buy an overlord to begin with, it later produces two at once, far overshooting the population cap the strategy requires but allowing it to squeeze those 2 extra roaches into the rush. It’s all counter-intuitive, but the result is inarguable. Louis says that “This is the type of non-obvious optimization that genetic algorithms excel at.”

Truly, we are living in the future. Here’s a tutorial on how to do the 7 roach rush.