Speed up budget optimizer

[AlfredoJF]

Is there a recommended way to speed up the budget optimizer within pymc-marketing?

I know that providing good initial guesses is beneficial, but has anyone else done something in this space besides that?

[wd60622]

Hi @AlfredoJF ,

Good question. There is this thread: #780. I suspect the use of grad in scipy optimizer will have some improvements.

The initial guess kwargs should be available in the minimize_kwargs. I believe that was introduced in 0.10.0. That seems to be the x0 kwarg

Carlos is currently working on this PR #1151 which will make some changes which might include a speed up. Feel free to comment in the PR if you have thoughts

[AlfredoJF]

Thanks @wd60622!

Unfortunately, x0 is not part of minimize_kwargs but still I think it's ok for now.
https://github.com/pymc-labs/pymc-marketing/blob/fa44cc7fbd134c8a3aa302d0bf2687383f706bff/pymc_marketing/mmm/budget_optimizer.py#L193C9-L193C76

What about setting different params for minimize_kwargs? I'm curious about setting ftolat that level.
https://github.com/pymc-labs/pymc-marketing/blob/fa44cc7fbd134c8a3aa302d0bf2687383f706bff/pymc_marketing/mmm/budget_optimizer.py#L204C1-L207C14

Might these settings improve the processing time without compromising the allocation? Have we tried before?

minimize_kwargs = {
    "method": "SLSQP",
    "options": {"ftol": 1e-6, "maxiter": 500},
}

@carlosagostini Keen to explore this in the meantime grad and PR #1151 are merged.

[wd60622]

Makes sense. I overlooked that then.

Would you be interested in making a PR @AlfredoJF . I see you raised a similar issue before: #876