This is cool news. Partly because they finally seem intent on fixing their cheating with superhuman micro, but mostly because repeat games could allow humans to develop counter strategies. The exhibition matches earlier this year only had the humans play one match against any given bot, thus giving the bot a huge informational advantage.1The neural nets were trained through supervised learning on human matches before being pitted against each other. These new matches will hopefully end up being more fair. However, Blizzard’s press release is vague enough that Deepmind could still decide to play only a hand full of games, which would prevent humans from gaining enough knowledge to devise counter strategies. Worse, the Alphastar team could decide to stop playing once they observe online message boards sharing good tactics between human players or any other nasty hacking of p’s.
The hype labs haven’t been doing well in fair play or reproducibility so far, but I’m willing to hope they’ll improve.