How does one go about injecting equations into a specific population?

[Hasith2005]

Hi @MilesCranmer ! Before I say anything, I am completely new to Julia. Here's the question: if I wanted pause evolution at a certain point during the loop and inject my own equations into a specific population based on certain conditions, how would one go about doing that? I believe I must edit the backend but is there a starting point? For context, as a little project, I wanted to see if equations generated by an LLM which receives population-level stats (like the kind of equations being evolved, equations with the best and worst fitness etc. ) as well as general information about the domain of the dataset can help guide the search to terminate quicker or generate more accurate equations that fit the dataset as opposed to just pySR on its own.

Apologies if this is too vague. I can supply more information if needed! Thank you!

[MilesCranmer]

You can use the guesses argument to suggest guesses, see https://ai.damtp.cam.ac.uk/pysr/dev/api#determinism. Let me know if that helps

[Hasith2005]

Thanks for letting me know @MilesCranmer ! I think guesses is only for a nudge at the start, correct? I wanted to inject LLM output during search dynamically. Do I need to meddle with the backend for that to work ?

[Hasith2005]

hiya @MilesCranmer sorry for nudging but your input would be highly appreciated !

[MilesCranmer]

No worries. Yeah I think what you should do is use guesses with a smaller niterations and warm_start=True (which propagates the state forward - though not sure you even need that, you could just put all the previous state into guesses).

In other words you can do something like:

model = PySRRegressor(..., niterations=5, warm_start=True, guesses=[])
model.fit(X, y)
for _ in range(n_prompts):
    model.guesses = llm.make_new_guesses(model.equations_)
    model.fit(X, y)
    # log whatever you want each step

Also worth checking out https://github.com/trishullab/LibraryAugmentedSymbolicRegression.jl which builds on top of the SR.jl backend

[Hasith2005]

Thank you so much!!! I will take a look at that and the link as well :)