A Rudimentary AI for StarCraft Tasks

StarCraft is a complex game with both spatial and non-spatial information as well as temporal dependencies on that information. Even doing seemingly trivial tasks can prove to be frustratingly non-trivial as translating the input information to output actions isn’t simple.

Based on this paper by DeepMind, my project (GitHub and write up) seeks to show that applying recurrence on the convolutional layers can work as opposed to treating the convolution and LSTM layers as separate. It also introduces a metric to ascertain how confident a policy-based agent is in its actions: uncertainty.