Starcraft AI blog

I only see a few points to note in this year’s win percentage over time graph.

Most bots don’t appear to have gained in win rate after round 50 or so (50 games against each opponent). It’s not surprising. If a bot’s learning model is a k-armed bandit with a modest k, which is common, then the learning algorithm will have saturated by then. The curves up to that point are dominated by statistical noise and the struggle to adapt to opponents whose learning skills are at a similar level.

There are signs that #1 BananaBrain learned relative to #2 Stardust and came out ahead because of it. The wiggles in the graph are not entirely clear, though.

UAlbertaBot was able to learn a little bit beyond round 100. I suspect it learned more slowly because it plays random and has to learn separately for each race.

#4 Steamhammer was the champion learner. It passed #5 PurpleWave in round 149 and continued to improve very gradually. Steamhammer learns slowly but has a high ceiling because of its giant library of openings, and it thrives in long tournaments. Last year’s tournament was 149 rounds: If this year’s had been the same length, Steamhammer would have squeaked ahead in the very last round and its longer-term success would not have been clear.