Starcraft AI blog

In a game last month, Tyr by Simon Prins was winning until this happened.

Pro tip:Don’t do that. The Great Wall doesn’t keep barbarians out, it keeps you in. The protoss bot by Roman Danielis dropped a dark templar inside the base from a shuttle and cleaned house. (You thought that a dark templar carried a warp blade? This one had a mop.)

I shouldn’t pick on Tyr—I think Tyr is above average in placing its depots in tight clumps along the edge of its base. It hit a bug this time, and I think the bug has been fixed. Building placement is hard, and it’s especially hard if you are terran, because terran bases need a lot of room for production buildings and supply depots. Some bots space all buildings apart from each other so that there’s a gap between any two and blocking is impossible. Terran bots that do that end up running out of room and having to build depots and factories outside the base in vulnerable open areas. Here ICEbot spaced out the buildings in its main and its construction had to spill into the map.

The fancy solution is do a path analysis in deciding on building placement. Does it block units off from anywhere they need to get to? Can fresh units get to the front without threading a maze? The analysis will be trivial sometimes, but not when the base starts filling up. PerfectBot will analyze both its own and its opponent’s building placement, for both attack and defense. “Oh, the opponent built an obstacle course. The enemy army is here so I’ll drop there.”

Knowing when the opponent has built a wall is a valuable skill—notice how well LetaBot’s ramp wall-in works against many opponents. If you don’t want to do real pathfinding around buildings, you could try to recognize a wall when your units fail to make progress toward their goal. I’m not sure it’s easier and I’m pretty sure it works less well, but monitoring progress toward goals is a good idea regardless. Units that find themselves behind a wall may want to destroy a building in the wall, without caring whether the building is enemy or friendly. Tyr could have won if it had opened its own wall.

Most bots probably rely on heuristics to place buildings so that they rarely cause problems. Nothing wrong with that; it should work well enough and it will be easy, and there’s a lot to do so ease counts. But I hope a few authors will go for more ambitious solutions.