Starcraft Pro AI

Found this really nice article on a competition to make the best Starcraft AI.

Quote

The first thing our agent had to do was gather resources and produce buildings and units. This is both very easy and very hard. Well-known, well-studied opening strategies are popular in the StarCraft community, strategies that define production sequences for buildings and units. Like chess openings, they are designed to equip the player with a particular game configuration for transitioning to mid-game strategies. There are openings for early attacks with light units, aerial harassment, building up heavy forces for the mid- or late-game, and so on.

These openings are well optimized, and it’s tempting to simply choose one or several and hard-code the agent’s macro planning as a set of scripts or finite state machines. This is how StarCraft’s built-in AI and many other game AIs work: human expert knowledge is encoded as an explicit sequence of actions, with pre-defined transitions. Build tanks unless the enemy builds air units, then build anti-air units. If the enemy builds cloaked units, build detectors, and so on. The earliest version of our agent essentially used this approach, playing as the Terrans and massing tanks. Perhaps the first thing we learned was that this is not the best approach.

You can read the full 4 page article here:
Skynet meets the Swarm: how the Berkeley Overmind won the 2010 StarCraft AI competition

Shoutout to Meez and other gosu fags.