What happened at the 675000 mark?

[Stephen_B]

For the current run (RD53), what happened at about the 675,000 mark ?

[michaelgarvie]

This is actually quite complicated, but I will attempt to explain.

At 10000 (very early) a perfect fitness individual was found. However this was only detecting gate output faults. The next stage attempted to adapt this to also detect gate input faults, which requires more processing.

The way this was done was by inverting the priorities of self-testing (ST) and fault-secureness (FS) (p0 and p1). This is because if a circuit is FS wrt. gate output faults it is also FS wrt. gate input faults and hence the fitness for this was already 1. The only attribute that needed optimising was the ST which is usually low priority. If priorities were left like this then it would've lost the FS nearly immediatly when improving ST. However this method didn't work very well and ST never got optimised. So we had FS=1 but ST incomplete.

At 675,000 I discovered a way of evaluating detection of gate input faults using the same amount of processing. So from now on all runs will detect gate input faults as well as gate output faults. At this moment I switched the priorities back to normal and evolution got going again. However now we have ST=1 and FS is nearly there. It looks frozen, but I've checked it just now and there's nothing wrong with it. Just your usual punctated equilibrium.

So basically the fitness function has been made so that an evaluation of a more complete diagnosis capacity is made of circuits with same amount of processing. Hopefully it'll find a solution soon so we can move onto the next benchmark.