GA returns most recent (rather than best) value? about evolutionary.jl HOT 6 CLOSED

Indeed, the current implementation mutates elite population, and this is unconventional to the common practice of elitism implementation. It's worth considering to copy elite individuals intact.

from evolutionary.jl.

Also ran into this issue with the DE() algorithm at #112 , guess with the GA the results from previous iteration (which contain the actual minimizer) can be overwritten by later iterations, possibly somewhere

for the GA.

from evolutionary.jl.

Would the test result at commit timholy/ScheduleMeetings.jl@1dae62f

Iter     Function value
     0               14
 * time: 0.00019502639770507812
     1               14
 * time: 0.018580913543701172
     2               14
 * time: 0.018917083740234375
     3               14
 * time: 0.01920604705810547
     4               14
 * time: 0.01949787139892578
     5               14
 * time: 0.019734859466552734
     6               14
 * time: 0.019989013671875
     7               14
 * time: 0.02025890350341797
     8               14
 * time: 0.020509958267211914
     9               14
 * time: 0.02074599266052246
    10               14
 * time: 0.020987987518310547
    11               14
 * time: 0.0212399959564209
    12               14
 * time: 0.021517038345336914
Evolutionary.minimum(result) = 14
f(perm0) = 14
Evolutionary.minimizer(result) = [1, 2, 3, 1, 2, 3]
perm0 = [1, 2, 3, 1, 2, 3]

in agreement with what you expect? Got the result above when changing the minimize logic here

with

if (minfit < state.fitness)
     state.fittest = offspring[fitidx]
     state.fitness = state.fitpop[fitidx]
end

to avoid the result being overwritten by later iterations when it's not actually the optimized minimizer

from evolutionary.jl.

In the main branch I've modified the test to be a little more challenging (that solution is optimal, but it's also the initial guess). You could try your code patch with the current test?

from evolutionary.jl.

You may want to play with a mutation rate (decrease) or population size (increase) for more stable performance. Plus, instead of the default tournament selector, you could use more aggressive susinv to drive a population of low fitness permutations up.

I also found that the best permutation is [2, 2, 3, 1, 1, 3] with the fitness value 21, and it isn't possible to get such permutation with your mutation function swap2blocks as it only swaps numbers within a block. So, the starting permutation [1, 2, 3, 1, 2, 3] appears to be an optimal choice.

from evolutionary.jl.

In #108 I raised the question of whether it would make sense to provide an option to preserve the best-so-far individual ("elite" individuals are still subject to mutation, and hence they can be degraded as well as improved). Perhaps a copy could be made before mutation, so that you'd be sure to avoid any worsening? I'd defer to @wildart who knows much more than I about such matters.

from evolutionary.jl.