Broken counter for successful mutations in Genome.mutate()

[HenrikMettler]

I believe there is a problem with counting the mutations in Genome.mutate()
The counter successful_mutations increases only for mutations in the output region (see genome.py L400, L395: https://github.com/Happy-Algorithms-League/hal-cgp/blob/master/cgp/genome.py#L400). This causes the mutation_factor to be multiplied by a factor ≃ n_hidden_genes / n_output_genes (a bit lower in practice, since some hidden genes might be mutated several times)

Suggestion 1: add successful_mutations += 1 on line 400. But this causes the new issue, that also mutations that result in the same gene value are counted as successful mutations.
Suggestion 2: add a second output to _mutate_hidden_region that indicates whether there was a real change in the value of the gene (success?) and count only if a real mutation happened.

(PS. point 2 might also be necessary in _mutate_output_region as additional control for a "real" mutation)

[jakobj]

Good catch @HenrikMettler 👏 👏

I think your first suggestion should be definitely implemented, regardless of the remaining issues. We should count a mutation if one was applied ;)

Regarding the second point, this requires a design decision, I don't think there's the perfect solution. I would suggest to only return from the function after a gene different from the previous one was picked, sort of while gene was not changed: mutate solution. Otherwise nothing in the dna has changed, and I would hardly count that a mutation. In general I think it would be nice if the return values of _mutate_hidden_region and _mutate_output_region to be logically similar, do you think that's possible?

I'm not sure I understand why the factor is n_hidden_genes / n_output_genes. Have you checked this numerically? Could be easily done by fixing the second point first and counting the actual number of mutations over a number of trials.

[mschmidt87]

It appears to me that this commit introduced the faulty behavior:
91b62d5

[jakobj]

oops, my bad, my bad 😳 thanks for finding this @mschmidt87 🚀