Weighting by a Utility Function
Of course we should not try to make the decision with the highest
expected (monetary) value, but the one with the highest expected utility.
The utility of money is mostly not linear with the amount of money
(or deficit). The utility-function may, for example, look like figure 11.
As this step may complicate the paper more than it helps the result,
weights according to the utility function are not introduced in our
example here.
Sequential analysis and stratification
So far there has been only two possible actions: accept or reject. In
many cases it is very advantageous to divide the sample space into three
regions: Acceptance, Rejection and Further Sampling, se figure 12. lt
can be shown that for some general cases the sequential technique may
have an expected average total sample size of only half the size we would
get with the two-possible-action-technique.
As we however have indirect measurements of our errors, these general
advantages are not fully valid (except when we have an on-line computer
— as the Analytical Plotter case). Further investigation on this topic has
possibilities of leading to results of practical value.
We should also remember the might-be posibility to divide the popu-
lation space into different strata. To maximise the precision of estimation
of the character of a population, we should construct strata so that their
averages are as different as possible and their variances are as small as
possible, and the sampling fraction in each stratum should be proportional
to the square root of "the variance divided by the cost of an observation"
in that stratum.
Final remarks — further developement
There are good grounds for doubting the practical significance of the
whole business of writing down profit functions (or drawing curves) and
finding points of zero partial derivates. Such devices are merely aids to
thinking, as every real world problem is enormously complex. We think
that the techniques suggested here will be of assistance in developing the
field in which the mentioned working group "Standard tests" is working.
One goal for such work could be the following. Group the instrument
157