This repository was archived by the owner on Feb 28, 2024. It is now read-only.
This repository was archived by the owner on Feb 28, 2024. It is now read-only.
Compute distance with one point normalized #589
Description
Setting:
- I have a space with only categorical variables (most of them containing integers admittedly)
- ask and tell API is used with an gp optimizer
The first n_initial_points-many points are evaluated as expected. After that i get the following error:
[ ... ]
File "/usr/local/lib/python3.6/site-packages/skopt/optimizer/optimizer.py", line 333, in ask
x = opt.ask()
File "/usr/local/lib/python3.6/site-packages/skopt/optimizer/optimizer.py", line 305, in ask
return self._ask()
File "/usr/local/lib/python3.6/site-packages/skopt/optimizer/optimizer.py", line 366, in _ask
for xi in self.Xi])
File "/usr/local/lib/python3.6/site-packages/skopt/optimizer/optimizer.py", line 366, in <listcomp>
for xi in self.Xi])
File "/usr/local/lib/python3.6/site-packages/skopt/space/space.py", line 721, in distance
distance += dim.distance(a, b)
File "/usr/local/lib/python3.6/site-packages/skopt/space/space.py", line 492, in distance
" the space, not {} and {}.".format(a, b))
RuntimeError: Can only compute distance for values within the space, not 1.0 and 1750.
1.0 is not a reasonable value for this parameter (and far outside the bounds of the parameter space of that variable). Seems like one dimension is normalized and the other is not.
I am aware that using categoricals is probably not the most elegant way here. But this should work, shouldn't it?
When i have mixed variable types the categoricals seem to work fine. With Integers i often have the problem that the points to be evaluted quickly focus in the extreme values of the range. Is there maybe another way to solve this?