...
Decide whether we should change filters: Compute sum-of-merit for next-best 150 observations in current filter, as opposed to most important ~100 observations in other bands. This imposes the filter change penalty properly.
Decide Then decide on the best order for observing most important fields. This is now just a question of optimizing the sequence of 150 pointings.
We also want to impose a look-ahead penalty, in case there is an upcoming opportunity (based on sky brightness or zenith angle) to do better on a given field. But we want to guard against having the scheduler drive to later observations when urgency is higher. So we should distinguish between observation merit and urgency factor, and paradoxically we should observe now if urgency factor is much much higher later on. So what seems to matter is the gradient in urgency traded against getting a better value of m5 later on.
It would seem we need to compute a better-if-observed-later merit penalty that depends on the gradient of the urgency factor. So how about a penalty factor of the form
penalty=max over a period tau of (merit later/merit now)*(urgency now/urgency later).
This favors subsequent observations of higher merit, unless urgency is much high at the later time.
...
Some references
LSST science book
...