Starcraft AI blog

Here are my implementation thoughts about opening timing data, as mentioned yesterday. I haven’t decided whether this is what I’ll do next, still thinking. I will at least do something similar eventually.

the data

1. Record timings for all of Steamhammer’s openings, in a static data file to be read at startup. The timings should include the time when each tech or production building finishes, plus the number of workers and the army size and composition at the end of the book line (meaning the units produced; some may have been lost), and maybe a few other things. Time resolution of one second or a few seconds is probably fine. Even so, timings will vary from game to game, so maybe the timings should give low and high values, or mean and variance, or something.

Another idea is to accept that new openings will be added and reject the work of timing them before they can be played. Keep a database of opening timings and update it after each game. That’s only safe on a server which plays one game at a time; for tournaments like AIIDE the database would have to be frozen and read-only.

2. For each game against a given opponent, record the earliest time that each enemy unit type is scouted, including buildings. Steamhammer already does this, with its “unit timing” skill (implemented using the skill kit). Also record the timing and army size and composition of the enemy’s first attack, or maybe its first few attacks, or maybe all of its major attacks. I’ll see what helps.

using the data

The data about the enemy can be used to recognize enemy builds earlier and more consistently. Most bots have a small repertoire of openings. As Steamhammer plays, it can check its scouting data for the current game for the closest matches among recorded games. If the records say that the enemy followed up the same way in all of the close matches, then the enemy strategy is predicted to that extent. You can see it as a kind of case-based reasoning: Find approximately matching cases and generalize from them.

We don’t have to fully predict the enemy strategy to react to it, we only need to know constraints on it. For example, if we’re going to add static defense (whether written into the opening or in reaction to the enemy army size), then we can check records of when the enemy first attacked: Don’t build sunks too early. If a clever enemy notices the vulnerability and attacks early, too bad, but then we have a new game record and will know for next time.

The main purpose of the opening timings is to choose openings. The records of enemy games tell us the range of enemy play. When choosing an opening at the start of the game, before any scouting information is available, we can try to pick one with unit mix and timings that counter the range of enemy play. One basic adaptation is to try to always be a little greedier than the enemy, to get ahead in economy (except when the enemy is too greedy, then we can rush). That’s a principled way to choose 5 hatch before pool versus Dragon. Another is, if the enemy prefers certain units, we can pick openings that produce counter units at the right time.

Of course the data can also be used to adapt openings when our first choice was not right for the enemy’s actual play. It will be a while before Steamhammer can do that with generality.