Starcraft AI blog

The CoG results file is troublesome, so I’m analyzing AIIDE first after all. The purpose of a plan, after all, is not to be executed, but to be changed; contact with the enemy and all that. This year’s AIIDE detailed_results.txt file was easy to read and interpret. I only needed a couple small changes from last year’s script.

Here is my version of the crosstable. It is identical to the official crosstable, so it doesn’t include any new information. If somebody wants it, I can also post my version of the per-map results, but that doesn’t include any new information either.

The race balance is not too interesting this year. In the crosstable, we see protoss at the top, zerg grouped in the middle, and mostly terran on the bottom, so we don’t need numbers to judge the race balance. But here’s the table anyway. The random row and the versus-random column are the least interesting of all, because UAlbertaBot was the only random player.

And here’s the breakdown of how each bot performed against each race. The most surprising results are that Steamhammer did poorly against these specific 2 zerg opponents, although ZvZ is in general its strongest matchup, and that weaker participants UAlbertaBot and BunkerBoxeR somehow scored higher against mighty protoss than against middling zerg. In the crosstable above, we can see the matchups which were responsible for the surprises.

We also see that XiaoYi annihilated terran opponents, even though it came in a little below Iron overall. Comparing to last year’s results, XiaoYi’s parent bot SAIDA wiped the floor with terrans even harder.

Next: The voluminous per-map crosstables.