[cadynce] Meditations on Learning and a Big Back End

[Cross, Joseph]

Patrick -

Okay, very cool. We're down to touching up the face powder on our Miss
America contestant.

> We are talking about how to deploy a configuration that have 
> never been tested in the lab. 

Right. And looked at another way, we're talking about how to certify a
configuration that has never been tested.

> I conjecture that there is no way to do that with anything other than
a 
> statistical assurance. The goal is to first give them a 1000 tested 
> configurations and use the novel generation method as a back up when 
> none of them fit.

If we're being Bayesian, where probability is used to express strength
of belief, then you're saying that we'll never achieve 100% certainty
that our system will work. Sure. 

But yea verily we never had 100% certainty and we never will. (Drat!
There goes another one of those pesky alpha particles!) 

If we say to the Navy "Our process will deploy configurations for which
we have only statistical assurance of correctness," they'll hear "...
configurations that aren't as reliable as your old-school certified
configurations." And that's not true. 

Old-school certification provides only statistical assurance of
correctness. And it would be boorish of us to rub our customers' noses
in that fact.

We're talking public relations here, not technology.

Obviously we must not oversell our product. But I believe that we can
honestly say that we're going to provide mechanisms that will enable the
certification of untested configurations. (Then the means by which we
chose one of these certified configurations at run-time is an
engineering detail.) If somebody asks whether the cert board will have
to loosen its standards to accept our configurations, we say no.

In this way we avoid emphasizing the
statistical/not-mathematically-certain nature of the entire
certification process, with or without CADynCE.

- Joe