Genetics

Why is it that biological genes produce such an incredible spectrum of complexity and form the basis for life itself, but genetic algorithms are kinda mid? Maybe there's more applications that I don't know about, but my understanding of genetic algorithms is that they're occasionally used for optimizing parameters in high-dimensional parameter search spaces, but don't seem to compete with state of the art gradient descent based algorithms. I've seen papers showing genetic algorithms can be effective in reinforcement learning when the path towards optimal control isn't so clear, and I'm sure there are other applications that I'm not aware of. I honestly don't know that much about this field but I'm gonna yap about my opinions anyway because I want to. Yet genetic algorithms clearly don't live up to their biological namesakes. How can we build a new class of genetic algorithms that allow for the kind of computational paradigm shift that biological genes invoked for life?

Genetic algorithms today rely on heavy implicit bias, and while they can be effective at exploring certain kinds of parameter spaces, they seem strictly bounded by the algorithms they optimize. Tree-based genetic algorithms build expressions out of mathematical operations, but seem to produce brittle structures as any kind of tree modification results in a vastly different computation. Lisp-based algorithms and other code-based GAs have completely lost popularity after showing much promise that seems to have been limited to the academic.

Biological genetics function in a much freer way that allow for huge amounts of emergent complexity. Genetic algorithms have genes which code for explicit functions predetermined by the programmer, and their evolution involves recombination of these functional units. The functional units then interact in a procedural fashion based on the current computational paradigm. Biological genetics have genes which instead code for building blocks that allow for limitless amounts of novel structures to be generated. These structures then interact en masse, producing another layer of emergent complexity while allowing for much more flexibility in the way that these functional units interact.

Something about how genes produce proteins that regulate genes, allowing for feedback mechanisms. Something about all the different levels that genes and their products can interact with each other and how that allows for more layers of emergent complexity and greater exploration of possibilities.

I'd like to explore new paradigms for genetic algorithms that allow for more emergent complexity and more flexible interactions in the vein of biological genetics, while leveraging the speed and control which modern computers provide. Possibilities include:

Explicit simulation of biological genetics - advances in computational protein folding as well as molecular dynamics allow for detailed simulation of biological processes in silico. This process could most closely mirror the complexity of biological evolution, but is extremely computationally expensive and would require further development of simulation software and hardware to allow for anything on the level of biological evolution. This is the naive approach, and will probably function eventually but lacks flair.

Basic computer instructions - use instructions at the level of CPU operations or the like. Allows for universal computation while meshing nicely with current computational paradigms, so could be more efficient. However, still suffers from brittleness, and previous approaches (that I know of) in this direction have showed some academic results but haven't produced anything that seems truly useful. Could use more experimentation to understand the limits and potential.

Abstract computational systems - manipulation of universal computational systems that could theoretically produce arbitrarily complex behavior. Hard to define how these computations function and could require huge amounts of computational resources to produce anything practically viable. Allows for wholly new structures to be generated and tons of flexibility in the way computations are interpreted and applied. My preferred approach in terms of aesthetics and potential.

Honestly I just wanna play God I don't really care how