Notes on lpector-gptp10-preprint

Fwoosh, this is a relatively hard reading. Some of my thoughts:

Lots of work is being put into semi-random mutation of primitives, whether in the generational or the variational mechanisms of the programs. That makes sense, because this is, after all, genetic programming and not anything else. Maybe it’s because I’m taking a class with Lynn Miller or I’m just plain wrong, but I wonder if there shouldn’t be some sort of addition of protein programming? As in, modification/addition of code that is already known to have some sort of usefulness. Because it seems that although much energy is spent in creating functional programs, if almost everything was already functional, and there was just some sort of cherry-picking method or as already exists a tournament, lots of computational effort could be saved. Of course there would need to be research on what constitutes “useful” (which I assume GP is supposed to automatically create), but it’s just a thought.

ADD: I’m guessing the “code” operator/type of Push kind of functions as a sort of protein, in that it modularizes various primitives to create a (hopefully) functional piece of code. Evolution is supposed to find which atoms make the right code – perhaps there should be some sort of over-optimization process to facilitate this creation? There doesn’t need to be explicit optimization code, but I’m thinking more in terms of libraries/databases of “proteins” that work, or those that don’t. The big problem would be that it any such thing would theoretically be infinite and would probably slow the process down by a lot. No optimization to be gained there, but just another thought.

MORE ADD: The fitness process has to be run through each fresh run of a new environment, and better code is “saved” through the “mandatory improvement” implementation – interesting. Worth mentioning is the selection process:

  • Prefer reproductively competent parents
  • Prefer parents with non-stagnant lineages
  • Prefer parents with good problem-solving performance

Also, I was wondering whether there is already literature on the effect of programs on the problem they are trying to solve. As I see it, the problem is the “environment” in which the programs are supposed to evolve to match or die in. But it’s not just the environment that has an effect on the various organisms inhabiting it, but the organisms change the environment too. Maybe bringing ecology is too much, but it seems like if there is major change due to this, it should be accounted for.

One question I had is on the purpose of reproduction in real life. Obviously this is a question that nobody knows the answer to, but it seems like one side effect of a well-functioning reproductive system is that one will have more offspring. Do populations of species increase as their reproductive systems become easier to manage and thus work? How does this relate to autoconstructive programming?

Just one clarification needed – is AutoPush the descendant of Pushpop?

This is really interesting stuff – I wonder if there is anything to be gained from existing poly/metamorphic viral code and anti-viral (specifically heuristics) literature.

0 Responses to “Notes on lpector-gptp10-preprint”

  1. No Comments

Leave a Reply

You must login to post a comment.