Starcraft AI blog

Today I put in the workaround for yesterday’s visible vespene BWAPI bug. Steamhammer now considers the gas reserves of an enemy refinery to be inaccessible. At the same time I fixed a rare bug that could briefly give a gas supply of 0, handled mineral and gas amounts of 0 in a more informative way, added features to improve the debug display, and made unnecessary efficiency improvements (“the cost is negligible—but—but—look how much I can reduce it!”). It added up to refactoring most of the code, but there’s not much of it so it wasn’t long.

Some later day I’ll add tracking of the bot’s own mining. Any minerals or gas missing from the map that Steamhammer did not mine were mined by the enemy, putting one lower limit on the enemy’s total resource gathering. Another lower limit comes from counting enemy units. The first use of enemy mining information will be for scouting: Overlord thinks, “these minerals have been mined, I don’t need to fly any farther to know that the base has been taken.” Look at the map Python. An overlord moving from the base at 3 to the base at 12 will encounter the mineral patches first, and can see that they have been mined before it comes into the sight range of any worker doing the mining. Identifying the enemy base even seconds earlier is an advantage worth taking.

The enemy’s resources will be valuable for strategy calculations. The same overlord on Python can judge whether the enemy is playing a low-econ strategy even before it gets close enough to count the enemy workers. The total resources mined by the enemy, as long as we have it, tells us about the enemy plan. It’s possible to find a hard upper limit too, but it seems complicated to get an upper limit that’s tight enough to be useful. Workers can be made this fast, bases can be taken this fast, a mineral patch can be mined this fast, it’s not as simple as counting stuff on the map.

In the long run, I envision an enemy resource usage estimator that takes visible resource amounts as only one source of information. Saved game records in the opponent model would be another, and a library of known strategies might be another. It feeds to an opponent plan recognizer, which feeds to a strategy planner. In my mind, resource tracking is a step along the path to smart strategy adaptation.

Stepping back to take a wider view, there are two ways to be smart: You can learn stuff like CherryPi with its neural networks, or you can calculate stuff like FAP—you can use knowledge or you can use search. I think you should ideally use both, like AlphaZero, because they support each other. Calculated limits on the unknown value of the game state are good inputs to help a machine learning algorithm.