Learning from (to infer something outside
) is doomed if any 'unknown'
can happen.
If is large, we can probably infer unknown
by known
.
So, if
and
small ==>
small ==>
with respect to
.
Now, we can use 'historical records' (data) to verify 'one candidate formula' .
But in real learning, we have to deal with some BAD
sample: and
far away --can get worse when involving 'choice'.