Starcraft AI blog

When the map is mined out and the game is not over, the nature of play changes a lot. Bots rarely get that far, of course—almost never. I’ve been trying to think of a bot game I’ve seen where the map was mined out, and I can’t think of one. But I claim that the example is still interesting to bot authors, as least as a sign for the future. Bots will need more flexibility and reasoning ability when they improve and do reach the utter endgame.

Instead of “when the map is mined out,” I should say: When the players run out of resources. There may be plenty on the map, but resources in the ground don’t help when the players can’t mine due to lack of workers, or lack of a command center/nexus/hatchery to return resources to, or lack of air transport to move workers to and from the resources. Horang2 vs Jaehoon in 2012 is an example. I don’t know of any games where both players could mine except that the opponent prevents it, but it is theoretically possible. For example, both players could have dark templar (unable to fight each other without detection) at the last remaining minerals, so that if either player tries to mine, their probes will die. In a mixed example, if the last resources are on an island, one player may be unable to mine due to lack of transport, but could have defenders at the island to prevent the other player from mining too.

As resources wind down, pro players start to get rid of the workers that they no longer need, to free up supply for more military units. They may send SCVs charging into the enemy army to be annihilated. I’ve also seen probes gathered and stormed to death. By the time all minerals are mined, depending on how the game went there may be only a skeleton crew of workers to use up the last banked reserves in construction or repair.

Players will evaluate whether they have winning chances. A player who can’t win, or who is risk-averse, will strive for a fortress that the opponent cannot break. Whether aiming to win or to draw, players will switch toward low-resource unit mixes. A low-resource mix usually includes many spellcasters, especially vessels, queens, and dark archons, but also high templar with hallucination and storm. It also usually includes fighting units that are efficient for whatever reason, tanks (for range and power), lurkers or dark templar (especially if they can stay undetected), hit-and-run units like wraiths that can escape pursuit, and so on. Reavers may run out of scarabs when there are no minerals, but can be worth it if they’re left over from earlier in the game. The mix depends on the situation; if zerg has queens, protoss wants archons that are immune to broodling. Figuring out a good low-resource unit mix for a given endgame is a different skill than figuring out a midgame unit mix.

You need reasoning ability to do this 100% right. If you have only a few buildings left, you have to figure out how to avoid being eliminated, and that controls your whole game plan: The opponent has these possible attacks specifically (I know because I’ve seen their last few units), so I should array my forces like this to have the best chance to ward them off. It’s another case where you want explicit goals that you can reason about.