[squeak-dev] Neural based evolutive testing (was: The Trunk: Kernel-nice.1402.mcz)

Douglas Brebner kirtai+st at gmail.com
Wed May 19 00:27:15 UTC 2021

On 18/05/2021 20:44, Nicolas Cellier wrote:
> Hi Douglas,
> If the number of failing cases is one out of 1,000,000 or more, the
> probability to find one is infinitesimal.
> The percentage of failure in the sqrt case was even smaller (it
> requires crafting the trailing bits of integer whose highBit > 100)...
> I don't believe that fuzz testing is really the effective way in this context.
> Only the setup is effective (very simple and cheap - though we have to
> bound our LargePositiveInteger...).
> Shooting in the dark is effective only in crowded places.
> White box testing is certainly much more effective in such cases.

Oh yeah, in this context it may not work well. I was just pointing out 
that it can be useful.

Not to mention that even one in a million chances are likely to be 
smoked out with hundreds of millions (or more) fuzz tests.

More information about the Squeak-dev mailing list