Target function returns NaN

[bonh]

I have a complex target function that involves a nonlinear optimization under the hood. The optimizer might fail. What should the target function return to indicate to adaptive that this region of the parameter space is no good? np.nan ?

[basnijholt]

You probably need a custom loss function. Where if a NaN is returned, you set the region to 0 loss, meaning it won't be picked again.

Which learner are you using?

[bonh]

Okay, sounds good, I'll try that.

I'm using LearnerND for $R^3 \rightarrow R^1$.

[basnijholt]

OK, then try something like:

def custom_loss(simplex, values, value_scale):
    if any(np.isnan(v) for v in values):
        return 0.0
    return adaptive.learner.learnerND.default_loss(simplex, values, value_scale)

learner = adaptive.LearnerND(..., loss_per_simplex=custom_loss)

Note that I didn't test this 😅

[basnijholt]

Maybe you just want to make it less likely that something in that area gets chosen. You can do that with

def custom_loss(simplex, values, value_scale):
    loss = adaptive.learner.learnerND.default_loss(simplex, values, value_scale)
    if any(np.isnan(v) for v in values):
        return loss / 10
    return loss

learner = adaptive.LearnerND(..., loss_per_simplex=custom_loss)

[bonh]

That seems to work - thanks!