I suspect the same sort of deep-learning/reinforcement-learning that manages to play video games could be used to both assess and improve playability for the si

I suspect the same sort of deep-learning/reinforcement-learning that manages to play video games could be used to both assess and improve playability for the sight-impaired. For example, can a 'blind' agent, that's fed just the (stereo) soundtrack of a game, become good at it? (While not, add more audio cues.) Can another cooperative agent, which can see the screen but only emit audio, learn to help the blind-agent do better? (Train up this helper-agent during playtesting, then add its audio cues to the main game – either always, or in a special 'audio assist' mode.) (Parts of this could wind up like a rotated version of this "sound of pixels" work: http://sound-of-pixels.csail.mit.edu/ http://sound-of-pixels.csail.mit.edu/ But, instead of learning which pixel-regions are responsible for sounds, it'd be adding sounds to second-medium-communicate relevant visual information.)