Eugene Leitl wrote:
> I keep repeating this for umpteenth time: the whole idea behind the
> project is to circumvent the limitations of brittleware, by GA-finding
> a mutation function which mutates assembly without breaking it. If you
> have this Rosetta stone everything else is suddenly easy.
> We all agree that it would be much smarter to use robust systems, but
> that does not exactly help us to exploit the resources of tens of
> millions of machines existing out there on the net. Even now such
> resources are nonnegligeable, now consider the exponential growth of
> node numbers overlaid with Moore's law linear log plot, and think a
> decade ahead. There's gold in thar them nodes in Netland out there.
Sorry about that. I was primarity replying to Bryan Moss's scheme for winning the GIMPS prize, not your GA scheme.
I do notice, however, that your reply is basically of the form "if we had good, general-purpose code-writing GAs, then we could easily do all these things". Well, yes, I agree with you there - my point is that you can't pull off this kind of scheme until after you break the GA bottleneck, so it doesn't work as a method for attacking the bottleneck itself.
IMO you actually have two major constraints to deal with before you can get the GA systems you want. The mutation problem is the easy part - systems like the CAM-brain project seem to be well on the way to finding solutions via old-fashioned human inventiveness. The really intractible problem is how to write all those fitness functions. What you need is an automated means of writing fitness functions for any possible programming problem, whether it is a floating-point computation, a query optimizer or a network protocol stack. AFAIK no one has any idea how to even approach this problem, much less a viable avenue towards a solution.
Billy Brown, MCSE+I