Starcraft AI blog

I want to elaborate on something I passed over lightly yesterday. Making decisions in a complicated situation needs both search and knowledge, and you can trade off how much you rely on each. You can search few options with knowledgeable selection and evaluation of the options, or you can search many options with less understanding of each. The left column opposes the right column:

Humans are slow (neural operations in milliseconds) and have to act in real time, so humans rely heavily on knowledge for fast reactions. In a fast-moving Starcraft game, humans are primarily reactive and knowledge-driven (do what you practiced), and don’t have time to think through plans in detail.

Computers are fast (processor operations in nanoseconds), so computers should use more search than humans. (Also our learning algorithms are slower than human learning, at least in terms of data volume, so gathering knowledge is harder.) AI has a history of discovering how well search works. It makes sense.

Starcraft is real-time, so Starcraft bots should use more knowledge and less search than (say) chess or go programs that have time to think. On general principles, Starcraft bots should be more search-oriented than human Starcraft players, but more knowledge-oriented than chess programs.

Now consider the abstraction hierarchy, strategy, tactics, unit control. Unit control decisions happen the most often (since there are many units and each may need new orders frequently), so unit control has the strongest real-time constraint. Unit control, by the same general principles, should be as reactive and knowledge-based as possible, to free up time (allowing faster reactions and more time to think about tactics and strategy). Tactical and strategy decisions don’t have to be made as often and are based on a smaller amount of more abstract information, which favors searches that compare alternatives.

That doesn’t mean you have to do it that way, it’s just what the high-level view of the situation suggests. A strategy catalog (a knowledge base of strategies) makes perfect sense to me. Depending on what you count as “strategy” it may be that you can encode all your openings and counters and deceptions as knowledge and make strategy decisions mostly by lookup (where the lookup index includes an opponent model and dash of randomness). Or maybe not, I don’t know! Tactics seem more varied and call more loudly for comparison of alternatives by search and evaluation.