Starcraft AI blog

As LetaBot mentioned in a comment, CIG 2018 detailed results are out. They include a result file that I can analyze with my software, so expect the usual colorful crosstables in the coming days. For today, a few notes:

#7 ZZZKBot, whose basic strategy is 4 pool though it has added other strategies in recent years, was the only player with a plus score over #1 Locutus. On the one hand, it shows how dominant Locutus was. On the other hand, what did ZZZKBot do to win? Was it a prepared strategy? I’ll be looking into it. There are a few other interesting upsets.

The win rate over time graph shows which bots benefited from learning during the tournament, or at least which benefited from changes in play by themselves or their opponents (you could gain win rate over time if your opponents mislearn about you). To my eye, there seems to be a higher rate of curving lines than I remember from past years. I’ll look into that too.

#17 Steamhammer, with about 35% win rate, was the bot with the most crashes at 1553 out of 3250 games. That’s a much higher rate than I expected, 47%. I still do not know why the crashes never showed up in my test environment. On the upside, Steamhammer scored 66% in games where it did not crash, which suggests that it did approximately keep up with general progress over the last year. In any case, it’s another reminder that reliability is a top priority. The second place crashing bot was #27 Salsa at 35% crashes, and Salsa finished dead last. The third place crasher at 30% crashes was #15 TitanIron, which also finished unexpectedly low after high expectations among watchers who guessed it was a fork of Iron. Every other bot had a crash rate under 10%.

And there is source. I will look into the code of some participants to see how they tick.