By not having too large default priorities, the GA seems to be more stable.
Instead of inverting the relation induced by fitness using `(1 /)`, we use
`negate` to get more numerical stability (or in this case, less instability
introduced by floating point precision problems).