Result of the last circuit?

Hi Miguel,

Just wondering what was the final result for the last circuit that we were working on (the largest circuit evolved)?

Also, would it be possible to have a time stamp or date on the hall of fame circuits as to when they were discovered?

Thanks,

prok

The best BEECOUNT circuit diagnosed 80% of faults. This is quite good considering the fact that evolution was quite low level. However I'm not happy enough with this result to put it in the Hall of Fame. The predicted time for BEECOUNT to reach full diagnosis was 4 more months at the current processing rate so I thought I'd drop it for a simpler sequential circuit.

I've also made evolution slightly higher level now by using four input LUTs and latches (before these were not given to evolution, it had to make them by itself). This means that less radical solutions might be found, but they will be found before we're dead. So it looks like we may have a result for DK27 soon.

Hi Miguel,

For your Genetic Algorithm package, are there options for duplicating "genes" that may be allowed a higher mutation rate but not be expressed unless the difference in fitness functions over the last X generations falls below a threshold? Also, have you ever given any thought to allocating different mutation rates to specific areas of a "chromosome" by having an unequal probability distribution for cross-over sites? What about tossing in the best hand design as an individual occasionally?

prok

Its really hard to know wether such fancy GA stuff works better or not. Since to know if it works you've got to do loads of these runs to get a statistically significant result. By the time you've done loads of runs there's no point doing them anymore!

There are some GA tricks though, its all in the papers. But to answer your question, mutation rate and crossover prob are uniformly disitrubuted along the genotype. Haven't inserted the best HD so far, but I'll experiment with this soon.

Are you reading up about GAs?

I've run a few SGA's while designing a self-clustering algorithm. I really like this project because of your use of islands which was similar to an idea I had while reading up on evolutionary theory (S. J. Gould, punctuated equilibrium). Been reading up on the functional aspects of genetics (epigenetics, imprinting, active RNA, micro-RNA, transposons, methylization, etc.) and on information theory and coding. Trying to use/find a method based upon actual genetics that can functionally allow aspects like punctuated equilibrium to occur, especially when the rate of fitness function improvement plateaus (micro-RNA and methylization come to mind in the biological model).

You only have to look at the graph (now the LHS bit is a bit squashed) to see punctuated equilibrium at work. Parts of neutral drift and then two or three fitness jumps togather... You can see the same by looking at the topology map for long enough. Once a fitter individual "species" starts spreading, it takes over the map.
All this info collected by this project would be really interesting to biologists. If we only had nicer tools to view it..