Benchmarks & research
Hajime-Y's unofficial NanoJev extension trains ViZDoom's PredictPosition task with feedback across several game steps, then compares the result with supervised training.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The Japanese article reports roughly 19% success before training and 79% after, against about 69% for supervised training. These are the author's runs on PredictPosition. They do not measure general Doom performance or independently reproduce a Jev benchmark.
Compare the repository's PredictPosition runs under matching task and assistance settings. The extension studies learning across several game steps. The author's criticism concerns demos that supply the predicted outcome of each option; it does not show that every Jev game or Doom task receives that help.