Genetic Algorithm 2D Car Thingy

Re:Genetic Algorithm 2D Car Thingy by ColinWright on Friday November 20, 2015 @08:15AM (#10600985)

Do you speak from direct knowledge, or is this a deduction you're making? I can't find any clues in your profile or other contributions, so I don't know whether this is your area of expertise - in which case I would bow to it - or if it's something you've deduced from other reading. The reason I ask is because your assertion is at odds with my understanding.

My understanding is that most small mutations lead to no apparent change at all, hence the robustness. Some small mutations lead to catastrophic changes, as do almost all large mutations.

But some small mutations lead to noticeable changes in the phenotype, and those are the ones that then get selected against. Some of them are beneficial in the existing circumstances, some are not. But the point is that it's small mutations that lead to a comparatively large change in the phenotype.

And that's my point, which you echo:

  > ... in genetic algorithms there is usually
  > much less redundancy and neutrality in their
  > genomes, because they are designed to solve
  > problems, rather than simply survive.

Yes. Current "genetic algorithms" don't actually do anything like the systems they have been inspired by. I wonder if real progress would be better if the implementations were more faithful.

[ Parent | on HN ]

Re:Genetic Algorithm 2D Car Thingy by raverbashing on Friday November 20, 2015 @08:28AM (#10601031)

Here's the difference between GA and GP

In this case it's GA, so your genome directly maps to characteristics. So (let's say) chassis density is N bits. Mutate the MSB of that and the change will be big

GP on the other hand would "build a car". And what usually happens during evolution is that some aspects gain resistance to mutation through some form of redundancy. So instead of Chassis density being a number it might be a crazy math expression that is not affected by small changes.

Also a lot of "genes" have no effect on the final product. But having a bigger genome (in this way) effectively lowers mutation rates.

It is "junk DNA" but actually has a purpose

[ Parent | on HN ]

Re:Genetic Algorithm 2D Car Thingy by TeMPOraL on Friday November 20, 2015 @01:28PM (#10602881)

> Also a lot of "genes" have no effect on the final product. But having a bigger genome (in this way) effectively lowers mutation rates.

Are you talking about biology or programming here? Because in case of the former, I think we've already figured out that there's no "junk DNA", and that DNA itself is more complex than just a code listing. There's gene expression, i.e. feedback from environment that selectively enables and disables functionality described in parts of DNA. And we already know that bacteria have useful information encoded on all three window offsets, making DNA re-readable 3 times for different needs.

[ Parent | on HN ]

Re:Genetic Algorithm 2D Car Thingy by raverbashing on Friday November 20, 2015 @02:18PM (#10603209)

Programming

[ Parent | on HN ]

Re:Genetic Algorithm 2D Car Thingy by randomsearch on Saturday November 21, 2015 @09:23AM (#10606862)

> Do you speak from direct knowledge, or is this a deduction you're making?

I'd describe myself as an expert in evolutionary algorithms, yes, and I've worked alongside Biologists for a few years, but I wouldn't describe myself as an expert in Biological evolution.

I don't know if this is just a misunderstanding of semantics, as you seem to have echoed what I said.

> My understanding is that most small mutations lead to no apparent change at all, hence the robustness. Some small mutations lead to catastrophic changes, as do almost all large mutations.

This is correct, but previously you wrote "often" small changes have dramatic consequences, which is incorrect if "often" means "a lot of the time" or "most of the time".

[ Parent | on HN ]

Re:Genetic Algorithm 2D Car Thingy by Udik on Friday November 20, 2015 @09:08AM (#10601181)

"Apparent" and "noticeable" are terms that point to characteristics that we, as humans, are able to notice. A 1% more chance of survival is hardly noticeable by a human being without the tools of controlled observation and statistics, however the mutation that allows it can spread very quickly through a population in a few generations. Optimization is a matter of fine tuning, so the effect of mutations has to be small enough to allow it. Sometimes a slightly larger mutation can create such an advantage that a cascade a smaller mutations ensues to fully exploit its potential. However I'd say it's more he exception than the norm.

[ Parent | on HN ]