Interesting AI research here. With input data as direct game model data, an AI was trained to be able to beat human champions. A Google TPU cluster was required to train the AI, but only a single desktop GPU was required to execute the trained model. One particularly interesting thing was that superior response times and click rate was not required to execute a superior strategy. Also, when the AI performed with window visibility management, it performed almost just as well as when viewing the whole map simultaneously at all times. It appeared that the AI maintained a “virtual window of focus” as it was trained from humans.
20190125/https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/
So, how do you setup your own AlphaStar on a LAN? Unfortunately, it appears that the full information required for this isn’t available as a single open-source project. However, you might be able to find the bits and pieces online and reassemble it from there. I’ve found DeepMind’s GitHub. Apparently, the whole of the researchers working on this go by the name of the DeepMind website, it’s not just a news publishing website, unlike some of the others.
20190129/DuckDuckGo AlphaStar github
20190129/https://deepmind.com/research/open-source/
20190129/https://github.com/deepmind?utf8=✓&q=&type=&language=