วีดีโอ

It's great. definitely needs some editing.

Wait, how has this only 13 views?

F'ing skill issue.

Nice video, people still visiting it, maybe you do a little writeup of it somewhere too? Probably your non-coding genes is not analogous, or only a bit, to how it functions in actual DNA. Maybe instead of more dimension, have a linear gene again, but each with a position, and if the x,y is high/low you get four quadrants of that gene? Maybe something like that for rotation too? Gets rid of that N².. For practically doing this particular task not sure if this is best. But then maybe the general idea was the real question here anyway.

This video presents brilliant ideas and makes them approachable to layman - which is why the comment section is full of people who think they can do it better than you (but really have no idea what they are talking about).

you can remove a variable, if instead of if (mode) forward(forwardBy); else right(turnBy); you'd've used forward(forwardBy); right(turnBy); forwardBy = 0; turnBy = 0; or if (mode) forward(num); else right(num); instead. although this kind of optimization may be pointless.

such a great video!

I like how well the two different "chromosomes" used for this problem illustrate an evolutionary problem with real genes (in that it's a problem for the organisms, that is). Your first chromosome was designed simply, but it easily got caught on local maxima because small changes could easily devastate the organism's fitness. This is because the chromosome consisted of a series of commands which were followed to get the final result, which roughly parallels how real genes work. (There's another layer between gene and phenotype in organisms-specifically, protein function-but otherwise, it's a solid approximation.) On the other hand, your second chromosome design was tailor-made for the problem at hand, including a 1:1 map between genes and the problem space, loosely analogous to an organism which literally had separate sets of genes for growing each of its body parts (ie, LEGO Genetics). This makes it harder for the algorithm to be caught on local maxima, which not only dampens the effects of mutation but also allows more and larger mutations to enjoy their beneficial effects without incidentally destroying everything that evolutionary history has built for it. A world where genetics worked more like that, with genomes designed from the ground up to be safe and flexible and allowing for maximum evolution, would be a world with biology that _wasn't_ riddled with weird evolutionary flaws that would be too costly to fix. The idea of using an evolutionary algorithm to optimise (parts of) the chromosome design of your main evolutionary algorithm is also interesting. Obviously, there is no way that real animals with their clunky linear genes could evolve some way of optimizing their genetic structure without incidentally dying because all of their proteins are coming out wrong, but a genome which is designed differently potentially could. Funny how the Intelligent Design folks don't pick up on this sort of thing. If we were designed, we were designed pretty lazily.

"We'll be discussing in my next video" - Never posts another video, lol...

This is absolutely brilliantly explained. The mouse on screen and motion and the way you broke down your discussion and argument is incredible. After having watched many evolutionary and neural network videos, I can say that most are less than half this long and less than half as informed and clear.

You are party to the machine apocalypse! I hope you're happy with yourself!

Great and very instructive. i loved how you dealt with the rigidity of code with a 2D chromosom from which you derive code, it gave me lots of ideas ! thanks !

Wow, that is really cool! ...pretty sure 3D chromosomes is overkill, though. I have an idea: how about each gene is only allowed to be used once per drawing? What I'm saying is, for any given attempt, once the turtle reads a gene, it fades out. Then, if the turtle lands on it again, it could...how about just execute the last instruction again? Or... just keeps going one more space? Hmm...sounded a lot better until the very end there...doesn't even help with the "two squares" issue... ...you know, suddenly, 3D chromosomes don't seem like such a bad idea...

I have to be harsh about this one: LOGO corpse is long cold; can we please stop humping it already? Could this project be any more bland and basic than it already is?

nice!

ogm; you have added more dimensions, and there is some very high complexity. u r going to run out of nucleotides to make yur chromosomes.

I really like your concept of "junk DNA", I like how you speak about how to setup your genes and all. I played a bit with genetic algorithm and the really bad mutation is always my main problem. Most mutation or reproduction ends up being really bad compared to the parent, so my population rapidly converge towards my first "good" DNA, which is not that good really since it didn't evolve much. Most videos mostly talk about the algortihm itself without going into details, mostly not talking about what their genes looks like. Good video, thank you.

It's funny because I've been a programmer for about 15 years now and I remember doing the logo turtle in elementary school and I HATED it. Probably because it was finite, the magic of programming imo is that it has rules and it keep going forever. Logo turtle is more an hardcoded thing than a program, but I guess you have to start somewhere, but for me it totally failed to make me grasp programming.

Man, your faith in humanity is way too high, dont put your cellphone number on a youtube video :P

For the linear chromosome model, would some of the problems be fixed if each gene encodes both a move and a turn?

I think evolutionary algorithm should be renamed to creation algorithm since there is no such thing as evolution without creationism. the creator had to input data for it to do anything at all.

Your linear chromosomes worked perfectly fine. Unfortunately the purpose of natural selection is to create organisms that survive, not produce perfect organisms. This is why you could only achieve command lines that were almost right, but never a perfect one.

No more videos? I was hoping to see the results of the experiment. Good work, though.

so.... almost 5 years later, where's that "future video"?

Well that locks like its gonna need a lot of calculation power why not go a simpler path start with linear DNA of length 2 add a mutation that adds DNA length force the turtle to draw each odd move and to not draw each even move. make the fitness function count the tiles the turtle has moved over by 50% and those untouched by 50% now count the correctly coloured tiles. that should be fairly simple ;)

The problem presented at 17:12 reminds me of the puzzles in SpaceChem

myBrain.exe stopped working!

To deal with the infinite loop problem, have you considered having two instructions per site, so that if a site is visited more than once it reads the second instruction the second time? To me it seems that this would avoid infinite loops. It would require a rather simple procedure to reconstruct the LOGO instructions regardless of this added complexity. Loved the video though, you make me want to change field :)

while having mathematical merit, you will never be able to optimize a system by higher abstraction planes, as they always use more performance. the way to go to optmize a computer program is to run a genetic algoritm on machine code.

I'm 12, why the fuck am I here?

I think, one could also use a changing color in the Master bitmap and the pen, in order to prevent a local maximum.

what if each gene was a position, along with information of whether the turtle should draw in that move or not? That way, if you mutated a gene, it wouldn't have incremental changes to the result, as long as instead of using the crossover method to pass down the parent's information, you used random genes in random positions from each parent. I'm a noob so i might not know what i'm talking about but it's an idea. If having few moves was part of the fitness function, you could also have the advantage of generating algorithms with a low memory footprint. If a you had an "off chance" of adding a gene with random information to a child, you could have an unlimited number of moves. I'm not sure about anything anyway.

Regarding the problem of performing recombination of 2D and 3D chromosomes, perhaps you could implement indexing with a Hilbert space-filling curve or a similar space-filling curve to map the genes to a linear chromosome. Then you could simply use linear recombination. Many space-filling curves preserve the spatial (sequential) relationships of their indices (in this case, genes).

Instead of encoding more in the genes to battle that 'dramatic' effect of small gene changes, why not rework the fitness function? Those two zigzag paths ARE very similar if you evaluate similarity based on progressively matching parts of the image under rotations and translations. The less parts are required the better the fitness. You could also try breaking the problem down into steps where the first step involves genetic generation of instructions to match at least part of the image and the second step requires stitching together the results of the first problem, genetically or otherwise.

Amazing!

And i thought i am smart... i hardly got it! ^_^

First video I ever found really explaining evolution in computers. I had to pause the video at 18 min. I got so deep in thinking. I made my own program that simulates evolution. It works. It's briliant too. I'll eventually get into more complex AI. but ohhh man! Thank you for helping me on the first step.

Rather than using a predfined number of moves, why not allow the number of moves to evolve as well.

13:23 the turtle moved in the wrong direction.

I can remember using logo in school computer club back in the 80's. It was bundled with the apple IIe.

just a suggestion, why not try adding to a pixel's red, green or blue component to a maximum determined by the pen down (eg. if less than 127, add 16 to red) and either play the Dawkins bimorphs style artificial aesthetics OR try match the Mona Lisa ? etc. awesome presentation, I love this field. thanks heaps.

Its beautiful.

Why not use another movement type where the turtle jumps? That way there is still only a 2D space regarding the genes.

Where can I find the relating videos you refer to? I would very much like to watch them!

This was dope af. Good job!

Was this video narrated by a voice generating software developed through evolutionary algorithms similar to this method?

Maybe you could select every single gene as a number (or array), made by 3 numbers combined (pen status, rotation and line lenght), treated as a single number, example: 1 bit for pen status, 9 bit without sign for rotation, 6 bit without sign for lenght = total of 2 bytes of data these 3 values might be combined as one number. You can then generate a number of random genes to build your chromosoms, and build your base population. This could be considered as a "basic move". I will follow this path, maybe also adding 2 coordinates for the starting point. an array of these 5 numbers should form my single gene, a chromosom made by M of these genes randomly chosen, a population of N chromosoms could then apply for the fitting function. Depending on the size of each of the 5 parameters you could have a usable number of combinations, as an example, choosing 1 byte for each coordinate, supposing to add the 3 dimensions provided above, your single gene should have a maximum of (2^32 -1) 4294967295 different combinations of values. You can even limit more the number of combinations by limiting rotation and lenght values or decide to use 2 complement to generate negative values. The crossover from two chromosomes in a population, should be performed by swapping the values of these 4 bytes together from one chromosome to another and create new chromosomes. Than you could apply some randomic genetic alteration on single bits of the gene numbers. Does it make sense? Sorry for my poor English. Kind regards.

17:31 wuthering about 4D XD

Benjamin, what did you use to create graphics?

Gave phone number and email on a youtube video. RIP