Benchmarks & research
Hajime-Y's unofficial NanoJev extension uses feedback over several ViZDoom steps to train the PredictPosition task, which the author compares with supervised training.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The Japanese article reports an improvement from roughly 19% to 79% on its PredictPosition task, compared with about 69% from supervised training. Those are the author's runs of an unofficial extension, not general Doom performance or a reproduced Jev benchmark.
Compare the repository's PredictPosition runs using the same task and assistance settings. The extension studies learning from several game steps. The author's criticism concerns demonstrations supplied with each option's predicted outcome; it does not establish that every Jev game or every Doom task uses that assistance.