"Improved" genetics can't help these snakes
Вставка
- Опубліковано 1 лип 2024
- All info is in the description!
0:00 Intro
0:11 Typical Genetic Algorithm Explanation
0:30 Programmers don't understand biology
1:10 How it works
3:09 Is it any good?
The snakes turned out pretty dumb, but I can't blame them, if I was a simplified genetic algorithm experiment, I'd take the easy route as well.
Genetic Algorithms Explained By Example - Kie Codes
• Genetic Algorithms Exp...
A paper about this: www.researchgate.net/publicat...
Music:
"Book Bag" - E's Jammy Jams, from the UA-cam Audio Library
"Cover Charge" - TrackTribe, from the UA-cam Audio Library
"February" - Vibe Mountain, from the UA-cam Audio Library
Intro sound effects from: freesound.org/
Haploid vs. Diploid evolution. Please make more videos on this subject.
Also, nice that you published negative results. They are under reported.
Okay, why doesn't this have more views?
Probably the aLgOrItHm
This channel has great potential. It will take some time, but keep it on and the views will come. 😁
I do believe the algorithm bout to blow this channel up
Don't know but UA-cam just recommended this today. Here before this video blows
Because tards keep commenting WhY DoESnT THiS HavE moRe ViEwS
I think one usecase for recessive genes would be if you want a population with individuals specialized for different things, like some ant spieces, and to some degree humans. Though might still not be the most computationally efficient.
very interesting theme
Got this recommended after 2 years, pretty interesting
same
good video! might try to implement this...
Wow. What are the chances that I randomly stumble upon a topic of my current research paper? (I first found out about your channel through the 10-letter word square video)
I'm currently experimenting with range expansions of diploid organisms. I'd like to clarify some points in the video:
1) I'll start with the scariest thing: you're basically using FOUR alleles in your simulation! Alleles are variations of genes - in your case "d0" and "r1". If we're talking humans, nearly all of our genes feature two alleles (more lead to complexity, and are unfavoured). So adding two alleles for a 'null' mutation, and another two alleles for an 'actual' mutation does not conceptually achieve what you intended. I believe, adding other loci (places for genes) with different effects is one approach (but depends on the model)
2) Even if we assume the four-allele genes, the reason the results don't differ much from the 'biallelic' case (where you would have three combinations: aa, Aa, AA) is that you limited the output to 0 and 1, in which case the only different outcome is "r1 r1 -> 1", which is the only combination that doesn't feature a dominant allele AND results in a 1.
3) A lot of the research I've seen does use diploid organisms, along with the most important functionality - genetic recombination (see e.g. "Expansion load: recessive mutations and the role of standing genetic variation", Peischl, 2015). That's the thing that allows the children in your simulation to express new genotypes
Having said all that, I like the videos on your channel, and I like your inquisitiveness. Keep 'em coming!
Does crossover even serve a purpose in GA. I feel like mutation is almost always better, and the main purpose of crossover in the real world is for genetic variation (eg for disease resistance which wouldn't be a problem in a simulation)
I'm not a biologist though or particularly experienced in GA just a thought from things I've read
Mutation gets stuck in local optima to a far greater extent than mutations + crossovers. If you're just looking to converge to the local basin, mutation is better. (Of course, in that case line searches are _also_ better, so...)
The hope when doing crossovers is that you're working in a space where optimums in a few different dimensions at a time are correlated. Then when doing a crossover you have a decent chance of transplanting the entire strongly-correlated partial solution for piece A of the problem, into something else that has a partial solution for piece B of the problem.
There are cases where mutation+crossover does asymptotically better than straight mutation, and significantly so.
If you wish to play around with this a little more, consider the trivial case of a very high dimensional space where one is simply minimizing x1+ x2+...+xn, over [0, +1]^n, starting from a randomly-initialized vector. This toy example is actually tractable to exactly solve for both mutation-based and mutation+crossover-based solvers... and as it turns out the latter does far better.
That being said, most of the time solvers like COBYLA do far better in practice.
2:03
i still dont know what the fuck is he talking about
1:27
SNAGE
wow that's. Interesting, damn and the fact that it only has a thousand views, and only 8 comments is crazy
Two years later: have you discovered anything more?
Nice video
How does it work?
Oh wait you’re the blender camera guy
How is this 2 years old but only has 231 likes and 3.4K views?!?!?
0.48 haha i feel this
but yeah really cool idea
Proof googles algorithm doesn't work right here