Can a small model decide which button to press in a real DOOM game? We built one and ran it on ruOS. ViZDoom told the bot where the monster was on screen. The bot moved left or right to line up, then fired when the target was close enough to the center.
What we learned. The tuned rule and both RuFANN versions won all 48 test rounds. The simple rule was fastest for one game. RuFANN could store the same rule in a model file and evaluate many decisions together efficiently. The trained model’s small reward gain was inconclusive.
What was tested in the browser cog. We built the actual cog and verified that it returned images and accepted the buttons chosen by RuFANN from test inputs. In the gameplay benchmark, ViZDoom supplied the target positio