Detect When the Policy Is Out of Distribution

Experiment log · Scout at Meridian General Hospital

Experiment log, Meridian General Hospital, east corridor. Scout's tight policy was trained in the nominal simulator and works there every time. On the real floor it is slower to respond, and last Tuesday it clipped the pillar. Nobody noticed the change in its behaviour until the crash: its velocity readings had looked different for days.

You will write the monitor that would have noticed. Four small functions: describe the training readings by a mean and a covariance, score each new reading by how far it lies from that cloud in units of the cloud's own spread, choose a threshold from the training scores for a chosen false-alarm rate, and combine several scores into one alarm that hands Scout to the cautious policy. The page generates the logs, measures false alarms and detections on shifted logs, and replays the episodes with your alarms.