Starcraft AI blog

Reality is distracting me. While we wait for it to lose interest and return to its doghouse, here is a new kind of data.

First, as a reminder, here is the same old crosstable, colored by winning rate.

Here is a new table, colored identically by win rates, but labeled with game durations. In each cell, the top number is the median length of winning games for the bot on the left, and the bottom number is the median length of losing games. Keep in mind that, for example, Stardust has a lot of wins and few losses (as you can see above), so its losing game times are based on little data.

Looking down the overall column: Stardust wins quickly and loses slowly, which probably just means that it’s hard-hitting and difficult to beat. The terrans have long games, as you might expect, though McRave is not far behind. Dragon wins slowly and loses quickly, which might mean it is vulnerable to rushes (well, we know that’s true). WillyT has virtually equal win and loss durations. Among the zergs, aggressive Steamhammer wins faster than it loses, McRave is roughly equal while slower overall, and defensive Microwave wins more slowly than it loses.

Some of the inside cells have interesting numbers. The Stardust vs WillyT cell has Stardust winning in 17:27 and losing in 44:44, which must reflect the slow-creeping nature of WillyT’s tonk build. Another cell with unbalanced numbers is McRave versus UAlbertaBot. Another thing to look for is reversals: BananaBrain won quickly and lost slowly overall and against most opponents, but against Microwave it was the opposite; both players reversed their win-loss timing ratios.