Starcraft AI blog

I am... amazed by the weird play of WOPR Z, whose name says that it is a zerg version of WOPR by Soeren Klett. Except for the name and the fact that they’re both in Java, I don’t notice anything in common between the jars. WOPR Z is based on Atlantis. I have the impression that Atlantis is not a well-known framework, so maybe more people should check it out.

WOPR Z seems to play the same versus all opponents. It opens pool first, gas, expansion and some zerglings, then builds 5 sunken colonies for defense and goes into mass mutalisks. That’s the entire strategy.

WOPR Z never makes another hatchery, so it is larva limited for most of the game, and it makes far more drones than needed for 2 hatcheries. It also doesn’t know how to recover from losing a hatchery. It’s not a strong player.

The mutalisks stay in a tight ball, which is typically good play (though not always). The mutalisk ball drifts slowly around the map, as if it somehow had trouble holding position. When the flyers accidentally sight an enemy, they pounce. By then the mutalisks are often many, and the target may be demolished in no time. It’s strange to watch.

I imagine that WOPR Z is a first draft that will get more work. Macro hatcheries and expansions, recovering from losses, and scouting for the enemy seem like skills that might be left out of a first draft. On the other hand, Atlantis claims to support most of those skills out of the box. I’ll be watching for updates.