"The extreme always seems to make an impression." -- Christian Slater
"I think the highest and lowest points are the important ones. Anything else is just...in between" -- Jim Morrison
If the primary concern is understanding the central behavior of a distribution, methods like simple random sampling or stratified random sampling might be more appropriate. Extreme sampling methods are specifically designed for tail behavior.
An interesting feature of the geometry of MOO is that it is the extreme bounds that interest us the most. The vert ivies of a polygon surrodng a clound of options are where the response surface changes. At the outer most vertices contaon the pareto frotnier, i.e. al ppnts whose position cannot be altered without comparomising at least ne goal.
assuming x ==> y then extreme x's lead to exterme y. an extreme sampling assumes that x ==> y and that exterme values of x let us find exteme sampes of y.
note: support the reader douts this "extremem-is-best" assumption.
That line of thinking could result ina whole stream of interesting
resarech papers that expore an alternate
sampling policy. Here,
do a binary split on a line draw between two exteme points. An alterante
policy would be split that line