Python: Symbolic Regression

Scientific progress, especially in the physical sciences, is a story of hypothesis producing testable predictions that are then either confirmed or rejected by observations (i.e. data). Even in predictive modeling, we generally fit a given model to observed data. What if we could go the other way?

What if we could take the data, and find the equation that would most closely have produced the data that we observe? Symbolic regression offers us an opportunity to do just that. It searches the solution spaces of possible equations, by combining mathematical operators with functional forms in a somewhat random manner guided by evolutionary success (e.g. piecing together the most promising mathematical forms using genetic algorithms). In this way, the resulting equation is free from assumptions (e.g. assuming the model is linear, a-la linear regression), or biases about how the dependent variable is related to the independent variables, etc.

[johannescartus]

Very interesting video, but now I have some questions 🤔
In practice it is rarely the case that one has data that perfectly adheres to an analytic relationship of the variables. E.g., the measured data might be noisy and/or the underlying problem does not have a closed form solution. How well does this method perform if you, e.g., we’re to add some noise to your example data? Also, since it will always find some result, is there any way to tell when it finds something that is actually describing the underlying problem (as opposed to just finding a random formula that happens to fit the noise)?

[nothingness]

Thanks for the video. Really very useful for Physicists.

[rahulbpillai]

thank you for the video. Can you make a video on Mutigene Genetic Programming for regression problem in python

[santoshkhanal]

Really nice video. Could you please make a video using symbolic regression on real-world data such as AutoMPG or California House Price or abalone dataset ( small dataset) or something like that? Thank you!

[FreeMarketSwine]

Can this be used for optimization or reinforcement learning?

[zhuoxuanli]

When I use SymbolicRegression to fit my data, the final formula is always a constant. I don't know why :(

[DistortedV]

I'm waiting for the Neurips paper that says: Symbolic Regression is solved, P != NP