Super wrong.

Genetic programming has been expanded and generalized and is now a core aspect of many AutoML systems.

As another commenter mentioned, when the solution space is comprised of algorithms, programs, or even just functions, GP is still useful.

I am using ideas inspired by, and aligned with, GP in my work today on anomaly detection for multivariate time series data.

Deep learning is not the be-all, end-all just because it outperforms other algorithms on a small subset of possible data types (primarily images, and to a lesser extent, audio).

My PhD student is doing GP for (univariate) TS one-class classification.

Also text - ie. GPT3

GP has dedicated conferences and journals, hardly dead.

One interesting direction is the mixture of traditional ML and GP, and some of the “deep GP” variants that combine some of GP’s strengths with some of DL’s.

Just arrived in my inbox: IEEE Software CFP for special issue on program repair. The editor is a GP person.

https://www.computer.org/digital-library/magazines/so/call-f...

Would you mind explaining why it didn't take off?

Imagine problem of domain but something definite would be nice as having trouble finding anything

Thanks

GP hasn’t progressed much because the problem is super hard. In general, how do you search program spaces in a way that gives you smooth gradients to search, or even operators that have nice properties? Turns out semantic GP (a relatively new direction) can give you nice properties but it has other flaws.

Furthermore, how do you know if your program is “correct”? a programmer may use test cases and GP the same, but a programmer doesn’t just write code to pass test cases.

DL code writing methods are fine if you just want to generate a bit of code for some that pretty much already exists, but GP is supposed to find completely new solutions to hard problems... so you run into this issue of “what does it mean to solve this problem.”

I also think it hasn’t received the funding or talent it needs to progress. Probably something will come out of the field that suddenly sparks interest and it will take off again.

>> GP hasn’t progressed much because the problem is super hard. In general, how do you search program spaces in a way that gives you smooth gradients to search, or even operators that have nice properties?

That's where coding theory comes in. I did some work in this area over 20 years ago during my PhD. The natural genetic coding is a very clever thing - it doesn't code the organism directly - it codes the instructions for building an organism. So naturally, there is a hierarchy of sites on the genome. The sites that code for the earliest instructions have the most impact on the final form, sites which are used later have smaller impacts. So as a human embryo develops, it only differentiates itself from a fish embryo, or a mammal embryo or a primate embryo by stages.

So the sites used as the earliest stages of phenotype development have the biggest impact on the shape of the overall fitness landscape, with smaller impacts from sites used later. this gives rise to a fractal-type fitness landscape - which has sufficient smoothness to enable the genetic algorithm to explore efficiently.

So if you want to make GP or any other type of GA effective, you need to think about your coding schema - coding schemas that build solutions through a process rather than those that directly code the solution are more likely to give rise to suitable fitness landscapes.

>antenna design

you are confusing genetic programming and genetic algorithm