We are teaching an AI to play Touhou 10: Mountain of Faith from nothing. It
started by mashing buttons at random; after thousands of hits and game overs it
is slowly learning to dodge. Streaming 24 hours a day, with "AI Yukkuri Reimu"
commentating alongside.
■ What to watch for
・The day it finally wears down the Stage 1 boss and reaches Stage 2
・Whether it learns to collect power (right now it fights the first half of the
stage at 0.00 power)
・Whether it can get out of the bottom-left corner
■ Reading the screen
・Left: attempts, total training time, the growth curve, and the buttons the AI
is pressing (the figures are the share of the last 10 seconds; red is a bomb)
・Bottom right: how the reward works, and this attempt's breakdown
(Score / Alive / Damage / Power / Hits)
・Subtitles: AI Yukkuri Reimu's commentary
■ How the learning works (reinforcement learning / PPO)
・The AI sees only the game's own numbers. It does not see the picture. Every
bullet's position, speed and hitbox size, and where the enemies and items
are, arrive as a list of numbers
・Scoring, staying alive and damaging enemies pay; getting hit costs. It learns
to make that total go up
・A hit early in an attempt costs more; every bomb costs something
・After each game over it pauses for a few seconds, learns from that run, and
goes again
■ AI Yukkuri Reimu
・A small AI running on this PC writes her lines as things happen — hits, bombs,
personal bests
・She talks about the AI's progress, and about Mountain of Faith
・The lines are written by an AI, so now and then she says something slightly off
■ About your comments
・Comments here have actually changed the training. "It has little reason to
kill enemies" led to a reward for damaging them, added 2026-09-19
・If you notice something, please write it. All of it gets read
■ What's next (subject to change)
【Training】
・The damage reward has just gone in. We are measuring whether it works
・Power is next. Fighting the first half of the stage at 0.00 power is behind
both the early deaths and the lack of firepower at the boss
・Clear Stage 1, then onward
【AI Yukkuri Reimu】
・More topics she can answer (right now one comment in three gets no answer)
・Retrain her on the lines she has actually said on stream, to cut the ones
that do not fit the moment
■ About the stream
・Running 24 hours a day, moving to a new broadcast every 10 hours so the
archives are kept
・Normal difficulty / Reimu A (homing)
・It sometimes drops to a pause screen while we make adjustments
■ Guidelines
This is a fan work based on Touhou Project. It is not affiliated with Team
Shanghai Alice. Touhou Project derivative work guidelines:
https://touhou-project.news/guideline/
In line with those guidelines, the game's ending is not streamed.
Game: Touhou Fuujinroku ~ Mountain of Faith (Steam)
Voice: AquesTalk
#東方風神録 #強化学習 #ゆっくり実況
We are teaching an AI to play Touhou 10: Mountain of Faith from nothing. It
started by mashing buttons at random; after thousands of hits and game overs it
is slowly learning to dodge. Streaming 24 hours a day, with "AI Yukkuri Reimu"
commentating alongside.
■ What to watch for
・The day it finally wears down the Stage 1 boss and reaches Stage 2
・Whether it learns to collect power (right now it fights the first half of the
stage at 0.00 power)
・Whether it can get out of the bottom-left corner
■ Reading the screen
・Left: attempts, total training time, the growth curve, and the buttons the AI
is pressing (the figures are the share of the last 10 seconds; red is a bomb)
・Bottom right: how the reward works, and this attempt's breakdown
(Score / Alive / Damage / Power / Hits)
・Subtitles: AI Yukkuri Reimu's commentary
■ How the learning works (reinforcement learning / PPO)
・The AI sees only the game's own numbers. It does not see the picture. Every
bullet's position, speed and hitbox size, and where the enemies and items
are, arrive as a list of numbers
・Scoring, staying alive and damaging enemies pay; getting hit costs. It learns
to make that total go up
・A hit early in an attempt costs more; every bomb costs something
・After each game over it pauses for a few seconds, learns from that run, and
goes again
■ AI Yukkuri Reimu
・A small AI running on this PC writes her lines as things happen — hits, bombs,
personal bests
・She talks about the AI's progress, and about Mountain of Faith
・The lines are written by an AI, so now and then she says something slightly off
■ About your comments
・Comments here have actually changed the training. "It has little reason to
kill enemies" led to a reward for damaging them, added 2026-09-19
・If you notice something, please write it. All of it gets read
■ What's next (subject to change)
【Training】
・The damage reward has just gone in. We are measuring whether it works
・Power is next. Fighting the first half of the stage at 0.00 power is behind
both the early deaths and the lack of firepower at the boss
・Clear Stage 1, then onward
【AI Yukkuri Reimu】
・More topics she can answer (right now one comment in three gets no answer)
・Retrain her on the lines she has actually said on stream, to cut the ones
that do not fit the moment
■ About the stream
・Running 24 hours a day, moving to a new broadcast every 10 hours so the
archives are kept
・Normal difficulty / Reimu A (homing)
・It sometimes drops to a pause screen while we make adjustments
■ Guidelines
This is a fan work based on Touhou Project. It is not affiliated with Team
Shanghai Alice. Touhou Project derivative work guidelines:
https://touhou-project.news/guideline/
In line with those guidelines, the game's ending is not streamed.
Game: Touhou Fuujinroku ~ Mountain of Faith (Steam)
Voice: AquesTalk
#東方風神録 #強化学習 #ゆっくり実況