Re: DifferentialEvolution: custom mutation and recombination functions?

On Wed, Aug 16, 2023, 22:23 Karen Róbertsdóttir <
[email protected]> wrote:

>
> Perfectly fine!  But, question:
>
> > The strategy_func would be
> > responsible for mutating (blending) members of the population together,
> > doing the crossover/recombination itself, and returning a trial vector
> with
> > shape (N,).
>
> So it wouldn't be told what other candidate to perform recombination with
> - it should pick recombination targets itself?  I mean, that's workable,
> just being clear on this.
>
>
The strategy function would be solely responsible for creating a trial
vector. It could do anything it wanted, so long as it returned a trial
vector the same shape as the problem description. Whether the strategy
function was sensible would be solely at the discretion of the user. The
fitness of the trial function is determined outside the strategy function.


> I can't speak for others, but I'm fine with receiving it in the [0, 1]
> range and then scaling it myself, to avoid the need for the stock functions
> to take that slight overhead hit.  But whatever your preference is works
> for me.
>

On further reflection it makes sense to supply the population in the [
bounds.lb, bounds.ub range]. It would be a copy of the population, so the
original array wasn't overwritten by the user.


(As a side note, Scipy's inability to save and resume the population during
> differential_evolution optimization used to be really annoying, given how
> long CFD optimization tasks take.  However, I did find a kind of cheap hack
> that I've been using ever since - since the random number generator is
> deterministic, I simply have the minimization function create a hash value
> for the candidate, and store the results of the simulation in a hash table,
> which I save to disk.  Then when I need to resume, I just load up the hash
> table, and if a candidate has been encountered before, it just
> immediately returns the previous run's simulation results rather than
> re-running the simulation. It's an awkward hack, and wouldn't work on tasks
> where the minimization function is really fast, but for slow tasks like
> CFD, it works  :)  )
>

If you use the DifferentialEvolutionSolver (warning, it's private and
subject to change), then you can step the solver very easily, it's an
iterator. For a single run through it should be very rare that the fitness
of a given vector is evaluated twice, only trial vectors are evaluated, and
they're always novel. If you want to stop/restart then I understand the
need to cache. TBH using very expensive objective functions doesn't sound
great for differential_evolution though, there's always a lot of function
evaluations.

Custom strategy_funcs can be of course used for things that have nothing to
> do with genes. For example: sometimes - as the docs note - a user may want
> part of their candidates' data to be interpreted as integer data. How do
> you mutate or crossbreed integers and have them make sense?  Well, that's
> really going to be task-dependent. Maybe the integer means "number of
> iterations" - if so, then perhaps simple interpolation is best.  But maybe
> it's a category - in that case, interpolation is incoherent, and you should
> either keep it the same or randomly pick a new category.  And if it is a
> category, and that category influences some other values in the candidate,
> then that may affect how you want to alter those values.  Maybe if
> round(candidate[0]) is Category == 3, then you want the floating point
> value at candidate[1] to be between 1.0 and 10.0, but if it's Category == 5
> then maybe you want candidate[1] to be between 1.0 and 5.0.  Again, it's
> task dependent.
>

`differential_evolution` already has an integrality keyword.

w.r.t implementation, it just needs someone to do the programming and write
tests. We always welcome new contributors for PRs.

A.

>

_______________________________________________
SciPy-Dev mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
Member address: [email protected]