Jev on SMAC: Can a Fast Choice Model Play StarCraft Micro?

What it does

The authors report zero wins in 35 official-map battles for per-unit actions in SMAClite simulation, not real StarCraft II. They also test group tactics and generated unseen battles against rules with the same information. Results were not independently reproduced and are not a like-for-like comparison with published reinforcement-learning scores.

How you can use it

Use this study's comparison design to examine a game controller: give Jev and simple rules the same unit positions and allowed moves, then compare wins. The authors tried both individual actions and group tactics in SMAClite, a simulation. Their reported results are not independently tested real StarCraft II gameplay.

Keep battles used to design rules separate from new generated battles used to test them. In this study, better control programs generalized better, while Jev did not add wins on top. That is a finding about these selected simulations and methods, not a universal verdict on model-controlled games.