Can the algorithms deal with invalid data, e.g. NaN or inf?¶
Yes. By default, whenever a simulation returns at least one NaN or inf, it is
completely excluded from the training data. In other words, the simulation is simply
discarded.
In cases where a very large fraction of simulations return NaN or inf, discarding many simulations can be wasteful. There are two options to deal with this: Either, you use the RestrictionEstimator to learn regions in parameter space that do not produce NaN or inf, see here. Alternatively, you can manually substitute the ‘invalid’ values with a reasonable replacement.
I.e., at the end of your simulation code, you search for invalid entries and replace
them with a floating point number. Importantly, in order for neural network training
work well, the floating point number should still be in a reasonable range, i.e. maybe a
few standard deviations outside of ‘good’ values.
If you are running multi-round SNPE, however, things can go fully wrong if invalid data are encountered. In that case, you will get the following warning
When invalid simulations are excluded, multi-round SNPE-C can leak into the regions where parameters led to invalid simulations. This can lead to poor results.