differential_evolution: make callback receive fun(xk)

[ZacDiggum]

In differential_evolution the callback function receives the current best parameter set xk and convergence but not the current function value fun(xk). For implementing a function value driven termination one needs to evaluate the function once more inside callback with the help of xk. Especially with small poplation sizes this creates an unnecessary overhead.
Could callback be altered so that it receives xk, fun(xk), convergence? Please?

[ZacDiggum]

see #6890

[rgommers]

Could callback be altered so that it receives xk, fun(xk), convergence?

This change would break any current usage of callback, because the function signatures would not match.

It is not easy to add an extra argument without breaking backwards compat, would need:

• the new argument to be at the end, rather than in the middle as you propose
• introspection of self.callback to check that it supports a second keyword argument with the right name

It is possible, and I can see the benefit, but the implementation would be ugly.

Any other opinions?

[dlax]

Any other opinions?

I'd rather have the "internal" iteration loop exposed in an API so that users would be free to invoke a "callback" as they want. DifferentialEvolutionSolver already abstracts out part of this in an iteration protocol (__next__ method), so it could just be one step further. We could even handle the solver initialization and termination using a context manager.

Something like:

with DifferentialEvolutionSolver(myfunc, mybounds) as solver:
   for i in range(maxiter):  # or for step in solver: # (if it is aware of maxiter)
        step = next(solver)
        mycallback(step.population, step.population_energies, step.convergence)
        if solver.converged:
            break

result = solver.result

Just a thought...

[andyfaff]

I'm not keen to change the callback signature because there are so many attributes that could be useful to expose. If they're all sent to the callback function then the callback function will get less and less like the callback functions for the other minimizers. I think the way that Denis displayed (i.e. using `next`) is much preferable as all the innards are available from the solver object. Indeed, the whole idea of __next__ was for the user to do something like this. Denis, I'm not quite sure what you meant by exposing the iteration loop in an API. Adding a context manager is a really good idea.

[dlax]

Andrew Nelson a écrit :

Denis, I'm not quite sure what you meant by exposing the iteration loop in an API.

I just meant what I illustrated in the example code ;) That would mean making the DifferentialEvolutionSolver part of the public API and have its `__next__` method return a "step" object instance with all information bits about current iteration status attached so that it can later be eventually extended without breaking bw compat. In general, it seems to me that with complex solvers like DE we should try to expose a reasonably flexible API rather than trying to make all use cases fit into the functional API with callbacks and friends. Giving control on the iteration loop is probably the first step towards this.

[ZacDiggum]

Hm, I'm giving up for now. But creating custom termination functions would still be useful...
I'm using differential_evolution as a robust solver for high dimensional model fitting, that's where the motivation comes from. So, if this topic could be solved, I'd be very grateful...

[ZacDiggum]

#6905 exposes func_val, nfev, nit and population to callback.
These could be used to create custom terminators.

[ZacDiggum]

#6907 now the diff is displayed correctly

[ZacDiggum]

@dlax's suggestion sounds good to me...

[dlax]

ZacDiggum a écrit :

@dlax <https://github.com/dlax>'s suggestion sounds good to me...

Okay, I'll try to scratch an initial implementation of this idea soon so we can discuss further.

[mdhaber]

It sounds like the preferred approach would be a different interface with more control (rather than adding more information to the callbacks), so I'll go ahead and close this. Feel free to reopen if I've missed something.