How to improve search?

[Gogex1]

My model seems to be stuck, as for 1400 iterations it didn't change hall of fame at all. What parameters can i change to improve it?

[MilesCranmer]

It really depends on the problem, but have a look through the options:
https://astroautomata.com/PySR/api/#pysrregressor-parameters

In particular the search space: https://astroautomata.com/PySR/api/#creating-the-search-space

and the constraints: https://astroautomata.com/PySR/api/#working-with-complexities

are going to be key to tuning.

The search size parameters are for making the search larger, which is also necessary for more exhaustive searches: https://astroautomata.com/PySR/api/#setting-the-search-size

[MilesCranmer]

Comments:

• Don't use neg since you already have - as a binary operator. The fewer operators the better.
• Don't use exp since you are already using ^ (same reason as above).
• Use / instead of inv since there are some simplification utilities available for /.
• I think the constraint on pow_abs is backwards; you actually want it to be (-1, 1) (any base, only constant exponents).
• Try setting up a constraint on * as well.
• Why tweak weight_randomize?

[DenisSvirin]

i thought, that if i increase weight_randomize, i will, potentially, find more suitable equation

[DenisSvirin]

Also if im going to change evaluation function ( and derivative). Is there anything else i need to change for proper algorithm performance?

[MilesCranmer]

There isn't a general answer for this; it is problem dependent and hard to know what values are optimal in advance. You could try to do a tuning with, e.g., GridSearchCV or wandb. Personally I use wandb for tuning parameters. Here is the sweep file I have used in the past: https://github.com/MilesCranmer/pysr_wandb/blob/master/sweep.yml where the performance is calculated with this function: https://github.com/MilesCranmer/pysr_wandb/blob/348964dcdc153ded6eb624481e8eeaf9d1e3f62f/pysr_wandb/pysr_wandb.py#L140 which takes the median over a variety of seeds and problems.

[MilesCranmer]

A good rule of thumb is to minimize the possible search space (https://astroautomata.com/PySR/api/#creating-the-search-space), use as many constraints as you can (https://astroautomata.com/PySR/api/#working-with-complexities), and maximize the search size (https://astroautomata.com/PySR/api/#setting-the-search-size). The other parameters you need to tune.