Starcraft AI blog

I did a quick investigation of the AIIDE bots whose names I didn’t know before. Some of them are interesting, and I will follow up.

#3 CSE and #4 BlueBlueSky were quickly identified as Locutus forks. The links are to diffs helpfully prepared by Bruce, author of Locutus.

#13 LastOrder is credited to Bilibili AI Research and is related to Overkill. Bilibili is a Chinese video sharing website, and their stock prospectus boasts of “our robust artificial intelligence empowered, interest-based content curation,” so it sort of makes sense. LastOrder uses the TensorFlow machine learning framework. The description says “all the production of unit (excluding overlord), building, upgrade, tech and trigger of attack are controlled by a pre-trained Tensorflow model using method similar to ape-X DQN. model are trained distributively against 20 different bots on a cluster of 1000 machines.” It promises more details later at Sijia Xu’s github (the author of Overkill). I see an 8MB file that looks like machine learning results. There is python code that I take to be TensorFlow glue, and C++ code that looks like Overkill. Overkill was strong in 2015 (which is the version still running on SSCAIT), but its learning experiments in 2016 and 2017 were not successful. LastOrder finished in the top half, so I gather that this time there was good progress, perhaps with the help of a larger team.

#15 MetaBot (link to github repo) is an evolved version of the familiar MegaBot. Instead of MegaBot’s 3 heads NUSBot, Skynet, and Xelnaga, MetaBot has AIUR, Skynet, and XIMP, a much stronger combination. MegaBot settled on Skynet versus most opponents, because Skynet is much stronger than NUS or Xelnaga; MetaBot should show more variety. MetaBot ranked higher than #20 XIMP or #22 AIUR (while Skynet did not compete), so there is at least a little evidence that 3 heads are better than 1.

#21 CDBot seems to have been forked from cpac, which is itself a not-very-deep fork of Steamhammer. It did not catch my interest.

#27 Hellbot is written in C# and relies on bwapi-mono-bridge2. It’s quite small and developed from scratch.