"StarCraft" is insanely popular for a reason. Developer Blizzard crafted what many consider to be the perfect real-time strategy experience, with three perfectly balanced factions waging war and a robust AI engine that is capable of effectively juggling the array of options available to human players. That's how most of us see it anyway. The AI is no match for the world's top players, and that's where the StarCraft AI Competition comes in. This year, a group of students at UC Berkeley took the top prize for giving top human contenders a non-human foe to sweat about with their ominously named Berkeley Overmind.
This is where I'd normally make a Skynet crack, only ArsTechnica beat me to it in the headline for a feature chronicling the birth of the Berkeley Overmind. "StarCraft" was released in 1998, but the possibilities for a user-created AI model -- "agent," as it is formally referred to -- only sprung up in 2009 when Blizzard released the Brood War API, a toolkit designed for that very purpose. The 2010 StarCraft AI Competition was the first of its kind, which is what makes the Berkeley win so noteworthy.
The "Overmind" in the AI's name is a reference to the game's Zerg faction, which is what the Berkeley team's creation is built to control. The lengthy ArsTechnica article -- which, it should be noted, provides plenty of detail to bring the uninitiated up to speed -- explains in details how the Zerg mutalisk units came to form the backbone of the AI's success.
Mutalisks are powerful units in the game for what you spend on them, but the advantage is balanced by the fact that they are weak to area attacks and difficult to control in large numbers. Since an AI-controlled faction doesn't have to worry about the complexity of grand-scale micromanagement, the programmers set their efforts to exploiting this advantage.
The Berkeley Overmind builds elaborate threat maps based on enemy unit sightings -- oh yes, it does recon too -- and knowledge of attack damage and distance specifications for each unit. Using that, it is able to plot safe paths for units to travel along, alter those same paths on the fly as situations demand and, in effect, harness the full potential of a mutalisk swarm in ways that no human player could ever hope to.
It's a fascinating breakdown, not just of how one user-created AI toppled a top human contender, but also just in terms of exploring the general thought processes that unfold as a game's artificial intelligence is constructed. Definitely worth a read.