The Smartest Creation! - Evolution - Evolving A Neural Brain! - Evolution Simulator
ฝัง
- เผยแพร่เมื่อ 16 ก.ย. 2024
- The Smartest Creation! - Evolution - Evolving A Neural Brain! - Evolution Simulator
😃 𝗦𝗨𝗕𝗦𝗖𝗥𝗜𝗕𝗘 ► bit.ly/SUB2ANTH... | ★ PREVIOUS EP! ► • Is Jurassic ARK Coming...
• Is Jurassic ARK Coming...
TURN ON NOTIFICATIONS BY CLICKING THE “BELL” BUTTON AFTER SUBSCRIBING!
Looking for A Server hosting service!? You should check out Nitrado in that case. They currently host all servers i have for Subscribers and it's the only service i'll use xD
You can check them out here: nitra.do/Antho...
- MORE ARK SURVIVAL EVOLVED VIDEOS! -
ARK 500 Giganotosaurus VS 100 Indominus Rex - • ARK 500 Giganotosaurus...
Totally Accurate Battle Simulator - • Totally Accurate Battl...
ARK Survival Evolved 80 Badass Spinosaurus - • ARK Survival Evolved 8...
20 TITANOSAUR VS 100 GIGANOTOSAURUS - • ARK - 20 TITANOSAUR VS...
500 T REX VS TITANOSAUR, ALPHA REX VS TITANOSAUR - • ARK - 500 T REX VS TIT...
GENESIS GOD OBLIVION VS LEVEL 30M DRAGON GOD & INDOMINUS - • ARK Survival Evolved |...
- FOLLOW ME ON SOCIAL MEDIA -
» Twitter ➜ / anthomniagame
» Facebook ➜ / anthomnia
Outro music: londonabove.ban...
WE MADE THE SMERTERST DENOSER!!
this neural stuff is complicated and now my real brain hurts from making a fake brain. Please send help in the form of advice comments.
Anthomnia it evolved into the smartest reptile alive
Anthomnia I love watching you play this
is it difficult to play God
life finds a way
Make a ring, next make feet along the outside of the ring, then add muscles to make the feat move, finally set it to run and watch.
One problem I see you making is that you are attaching too many muscles to places. if you attach one muscle to a joint or place that will pivot it will be able to either expand or contract. but if you attach 2 to it then you have 4 options for it to do and half of them make it not work. Either one muscle expands and the other contracts making the joint lock up, or you get them both to expand and contract at the same time making the joint actually work. The neural networks take time to figure out how to move muscles in unison so if you create a joint that needs multiple muscles to act together to function it will likely not get it to work. Look at how the bones are put together and think about what will pivot and make it so only one muscle controls each moving part. if you need one muscle to control 2 joints at once you can create a system to do that so that with how you place the muscle. Remember less is more with this kind of thing because it is easier to figure out how to control a simple system.
The neural network acts like this:
- The more layers you have, the more the creature will be able to do complex thinkings
- The more neurons/layer you have, the more situations the creature will be able to handle
- BUT the more total neurons amount you have, the more it will take to train!
This is the best explanation I've seen, noice
Wow that's deep
I actually winced a little when the dude said he was going to explain it to his understanding after saying it was going to learn faster-
What do you mean by complex thinking?
@@samuelromero5786 The creature will have more options/layers to choose from.
I just clicked on this video randomly at 1 in the morning and the ad that played was The Lego Movie. The entire movie. Best 1 hour and 44 minute ad I've ever sat through at 1 in the morning the day after Thanksgiving. Thank you TH-cam for finally getting ads right...
...for now.
OneTooth9272492 The whole movie, as an ad?
Hmmmm...
Whoever put that ad in is a legend
12:55 Neural networks are actually a lot more complex than that. If I understand the neural net from the game, each frame the network takes an input, does calculations on it (represented by the axons), then outputs a number which represents the strength of the muscle contraction. When you train the neural network, what is happening is that the network figures out how to do the calculations better.
He was just simplifying it
It's not really making better or more efficient calculations, it gives weight to the "decisions" or paths that produce better results towards the selected action; like jump, run, etc. If contracting muscle 1 and expanding muscle 2 makes the thing move backwards, the "brain" gives that path a negative weight. If doing that makes the thing move forward or towards the goal action, then that action path is given a positive weight. The next generation will ignore the negative paths and select the paths with the highest weights. The more often a path through the network produces a correct result, the more positive that path gets and the more often the brain selects that path.
3DPDK if each path is just a certain set of instructions for muscle contraction and expansion, then what is the difference between having one hidden layer with 100 neurons vs two hidden layers with 10 neurons each? Each have the same number of connections (or else am just bad at math).
Mozart 13, Your math is perfect, but a neural network works more like a statistical probability calculation. A single layer gives the brain 100 distinct choices, but what if none of those choices is completely adequate to accomplish the task. It chooses the one path that works the best, and then has no way to improve the final outcome after that. By having two layers of ten actions each, the brain selects the first best node and then combines it with the next ten actions, each in turn. It's also possible that a lower weighted first level choice might work better than the highest weighted choice with the best second layer node. That's where the "learning" comes in and why it takes many generations to build the weighted table. A 10 x 10 node matrix still results in 100 possible paths, but add another layer of ten nodes and the possible neural paths increase to 1000. It raises exponentially.
This is a really good video that shows how a "neural network" learns and also how it's choices can evolve over time:
th-cam.com/video/R9c-_neaxeU/w-d-xo.html
... and this network uses match sticks to do it's calculations.
A true Mozart very complex
At 7:40 the mind has separated itself from the body, it is no longer part of this plane of existence. It has become one with everything and everything with one.
Yeah, I thought it was like some 3rd dimensional being trying to escape its 2d plane. 👌
+Brent Porter 😂 That Was So Good 👍
I clicked the timestamp and I got pushed into an ad
Me 2
That's deep
The way you explained it... That's not how a neural network works! It doesn't choose a single straight path from the input to the output. All nodes are involved in the calculation.
7:30 When the brain is so smart it hacks the game
Skynet
are you from the past and just stole my joke?
15:00 "Dinosaur" twitching in place.
Me: You've ruined a perfectly good slug. Look at it, it has anxiety.
*kangaroo
SwaggyKawaiiPotato53 catzz: **SLUG. And I stand by that.
KorilD that's a roo
SLUG, i love slugs.
a
Imagine the pure, unbridled horror experienced by a human consciousness trapped in the body of a dysfunctional lump of sticks
David Knopp I have no mouth and I must scream
Devon Lock we call those vegetables
Sooooe? Stephen hawking?
@@Devonoton So middle school boys
I made a ball, it legitimately just rolls around
TheSpectreSoldier a youtuber tried too! They failed maybe you can tell him how, I forgot the name tho
Rccamo3 Maybe
Rccamo3 mattshea?
How did you configure the brain? How many generations did you need?
No matter how I do it, my rollers always get stuck in a bouncing hobble or lock up in a spasm. It's quite frustrating...
Same
Step 1, make a worm.
Step 2, put as many nodes as possible
Step 3, watch it take over the world
Yessss
Yessss
Yessss
Noooooooooo
Yessss
12:05 "Let me know your thots"
Now I want to see someone who understands neural networks real well play this, because I feel like there is more to the neural networks than have every layer have the same amount of inputs
gameridiotNOT me too. I've been trying to figure out neutral networks, but it's beyond me. I want to watch someone who knows how and explains it
I've seen Carykh make his own version of this game and play it years ago. He explained the mechanics of his really similar game really well and got me into programming somewhat. It seems like you guys are interested in that so look up his channel if you want :)
Look up The One and giant_neural_network, they even *made* their own takes on the concept
Unfortunately, the NN fails to learn effectively because changes to the motion alter the timings. You can't just kick blind, it needs sensors. I rewrote this in C and the NN had huge problems converging on a design - and really fit designs kept dying suddenly due to being unable to synchronise state in relation to the ground. Once I added two types of sensor the resultant designs were very stable and stronger lines didn't perish in a single generation.
This system is pretty much catastrophic. It's simply too chaotic for the NN to converge on a strong design. My C version with the added sensors resulted in creatures with almost 20 times the top speeds I could get out of the same structures without sensor data, and evolution was a far more positive influence.
Perhaps this isn't the best application of an NN/GA... you need to feed in data relevant to the task if you want it to converge sensibly.
Currently my best design on the C version is a forward leaning "hopper" ... it developed "arms" and learned to manipulate these in the air to always land slightly forward-leaning for the next hop... it's amazing that it never falls over. Not sure why it's faster than the running designs... maybe just because it's a more efficient motion and so it just always outlearns the more complex models,
@@garychap8384 lmao wtf. There is no reason to flex that you know C
Jeez, now we know why the great flood happened.
Anth took control of evolution.
you see Timmy's great grandfather once tipped forward and found out that this was a way to move a short distance at an extreme pace.
so now Timmy is currently laying on his face trying to find a way to move using only his ears.
You probably want to increase the number of layers based on how complex of an action you want. For example, han walking is very complex because it uses so many muscles at once. You probably want 6 layers for walking. For the number of nodes per layer, you want around 1 mode per muscle or something like 10(1.2)^x where x is the number of muscles. Don't quote me on that, it's just a guess based on my experience in game. However, be warned: more nodes = more time to evolve, because it needs to perfect more nodes.
Aryan Singh its our lord and savior
8:00 Creation has left the game
Your explanation of how the neural net works (at 14 mins) is rough as hell, but not far off.
I would suggest that for every muscle, you have the net be 1 extra wide and 1 extra deep. So for 10 muscles, 12 wide, 10 deep at each node. It's a massive over simplification (neural nets are a bitch) but i think it would be the best method
Oh boy here we go :)
Well at 5:30 ish, it's actually using the front leg for momentum (see how the body lifts slightly every fling, making it easier for the other leg to move the body), and the back leg is just jutting out as a counterweight, tilting the body up to the right, also helping the other leg move it.
I wrote a version of this in C, and what I learned is that the NN really struggles without sensors as it cannot determine proper timing... structure alone just isn't enough. After adding two types of sensors (one for ground contact and one for orientation) it learned SO much faster. Without sensors it may find a very efficient design for speed but it still cannot coordinate and often ends up kicking up into the air and landing on its back, and so the genetic line dies. The sensors allow it to learn to wait till the optimum moment to push... and this results in some incredibly quick designs, with all designs re-stabilising very quickly after changes.
So, that's the problem, and a bigger NN isn't likely to help much whilst it's always kicking blind and randomly getting lucky. The slightest change and the timing all goes out the window - then the fittest design just fails and stops propagating. I also find the mutations in this version are too strong and can swamp learned timing and kill lines fast... it's a bit too much luck involved in passing the fitness test.
Without sensors it's has too big a random factor per run to evolve well.
Still fun though, but it gets annoying once you realise the senseless (pun intended) brick walls it's hitting.
I dont read nothing
11:11 just picture a beetle gracefully galloping and leaping through the savannah
7:30 it crashed because the creature was so smart it has hacked his computer
7:40 this happens because there are two separate bone structures connected only by muscles, wich means the muscles have no limit and expand to infinite length
7:47 yeah, the brain is part of the neural system
When you started figuring out something about the neural network, you stumbled on a key weakness of the 'perceptron', the form of NN which has a bunch of inputs and a single output: there are many classes of problems it simply cannot solve. (It can't even act as an XOR: it can tell you if the A or B input is on, but can't tell you that one is on while the other is off.) This weakness, published in 1969, stalled NN research for years because people assumed it applied to multi-layer networks. Fortunately they were wrong.
Small numbers of neurons in another layer add MUCH more functionality than adding large numbers of neurons in one layer. Neurons work best when they are able to both activate *and* suppress other neurons. Everything from our hearts beating to locomotion to vision relies on this principle.
God isn’t lazy! He has a hard job too!
You do not know how neural networks work
Dan Doom this game doesnt know how muscles work
Felix White you have a avatar so your opinion does not matter.
well I know, and he doesn't know
Tenx wha....
PurpleMania lol
This game is cute, I don’t know why but it just makes me so happy.
Am I completely dumb or is it better (at least slightly) to put less outcomes with more inputs so it’ll compute the best outcomes with all possibilities. Or am I far off? I’m just thinking out loud.
No supports.
No overlapping limbs.
Can't comprehend that it is in 2D.
Unless you are trying to be funny, you are failing.
Make a ring, next make feet along the outside of the ring, then add muscles to make the feat move, finally set it to run and watch it go.
Michael MAnville Good points.
Cant spell feet.
Unless you are purposely mispelling it, which you have suceeded.
Please its hurting me
Mr. Plague God what are we doing with our live's. Lmao.
Mr. Plague good advice. Unless you are misspelling “me” your succeeding.
It’s hurting me
Joey Dragonz iz hurtin mehhhhh
Elian Dzemailov
Good addition!
You don’t know the difference between advise and advice
Unless you are purposely making that mistake, which you have succeeded.
It’s hurting too much!
trust me i'm an engineer ^^ i think you should have at least as many neurons on first layer as number of musculs or bones of the creature. my spider kind of crawled but more tests needed to be done :p
every neuron on first layer is your brain input, and last layer is the output/reaction.
Did your creature just learn flight? Without there even being air resistance dynamics? Awesome.
10 layers, 8 layers of 100 about 802 neurons, the biggest brain we have mapped is the Caenorhabditis elegans with about 302 Neurons, the next step up is the common Jellyfish with about 5,600 Neurons, a human has about 86 Billion Neurons so.... not quite the human brain
So the “brain” takes input, idk what the inputs are, but think of it is a number between 0 and 1, could be the hight of a joint or the angle between two bones or something a lot smarter, you can probably google it. All of these inputs (layer one) are connected to every point in the next one(layer two). How connected they are is randomized at first, but evolves. The connection is the function that takes a points value (for example 0,8) in the last layer and gives a value for a point in the next. Then all the points in layer two are connected to the ones in layer three, and the last layer is your output, values between 0 and 1 which is how strong each muscle contracts.
The best creature in the game is unironically just |~|
What evolutionary algorithms are actually attempting to do in a nutshell is numerically optimize an algorithm by slow modifications. So when you see that something is twitching a leg and not using it like in the case of the spider, what happened is that it reached a local minimum. Suppose you graph distance it could travel as a function of all the combinations of neural networks it could have. You'd have a surface plot of some kind like a multivariate function returning a height by position. The danger with such optimization is that depending mathematically on how you attempt to find this minimum maximum you basically do the equivalent of rolling a ball and seeing it fall into a hole. Except that hole is on an outcrop of rock overlooking an entire canyon. Until the ball gets out of that hole it won't get the extreme benefit from going into that canyon. In other words, there's a large optimization that hasn't been done yet because small tweaks to that value that would get the large optimization make it look like to the algorithm that it's actually worsening the AI.
Of course, that's from the standpoint that this thing can be looked at as some N dimensional function outputting a final distance where the function in this case is the creature you create and the parameters are the current state of the brain for the creature. It might not actually make much sense in practice to think of it as a bundle of numbers. Granted everything in a computer is a binary so in a weird way it works by default.
7:50 I got no strings to hold me down
that is not how neural network act.. its is not like choosing path inside it... its is more like computing values. Every nodes get their value by multiply(or other computing action?) different weights(which should be learned) with values from the nodes that are connected with on the last layer. so on and on the final output gets its value from multiply different weights and values from the last Layer. And it will decide on now situation this muscle should stretch or shrink.
P.S. sorry for my poor English if anyone feels hard of understanding the words or grammars...
I feel like the neural network works like this. The layers is how many variables you have which in this game is 2 for the most part, which muscle is moving, and then what the muscle is doing. So my theory is the first layer is the number of muscles you have and the second layer is how many options those muscles have which can be contract, expand, contract and expand repeatedly, etc. I believe that’s how it works.
Awesome anthomnia
A game like this, only in 3d, with combat being another goal for the creature. Like, still made of sticks and dots, but maybe with a model where you could place your creatures against other online like friends or just anybody. It would be I teresti g to say the least, like, have s function do exchange damage between creatures, and the ability to loose/kock off limbs from the creatures. Basically, this game but with combat involved, I can already think of all sorts of cool things to add for this, like claws and teeth all sorts of adaptations to make your creatures more deadly and tough enough to withstand damage. Even in 2d, this games would be awesome with combat. It's already loads of fun imo, I love the game. Just saying imagine evilving your creature to battle your friends creature they've been working on and evolving.
10:00 Hopping spiders are bad enough! *NOW YOU WANT HOPPING SLUGS!*
Lol
The neural network works by converting the inputs through activation functions, the output of the neuron is then sent along the axon and the value is multipled by the weight of the connection, the network learns by altering the strength of each connection, eg if this happens then move your left foot a lot and your right foot a little bit. Its a lot more complicated than you are explaining! especially back propagation. To create an optimal network you need to have a small(ish) network maybe 2-5 layers with an increasing level of neurons in each level and then a decrease down to 1 output: eg. 1 - 10 - 20 - 10 - 1 this will feed the outputs to converge on a simplistic output whilst also including a large space for computation in the centre of the brain.
Fun Fact:If you put a muscle between two bones parallel to each other,you can get 100% fitness.
The program needs to be set up so the creature always trys to move farther than the previous generation, and if it keeps repeating, it tries something completely different until it starts continual motion. Than it works on going faster and faster.
You probably can do that, to some degree. The trouble is in using any number of different shapes for your creations. You can tell one creature to do it one way, but any other creature has to use a different method. Machine learning is so that we don't tell it how, we tell it how to learn how.
Anthomia: (gives the creature an 9999 IQ brain)
The Creature: (gets dumber)
Nice Video
OMG YAY! It’s back! Last vid I was laughing so hard , I bet this one is going to be even funnier 🤩😂
ᗷᖇEᗪ- Cult Can I join your cult?
This game well I guess more of a simulator is pretty fun but a little more complex for me.
A big 1 bug of this evolution simulator is that whenever we create a creature that just stretch the muscles it shows 100% fitness 💪, so don't go to gym, tie both hands with one car and legs with another and just strecth😎
1:40: *It learns slower!* though it will be able to refine its movement to a greater degree than a smaller network.
Edit: Oh, 8 months ago... real useful advice, lol...
This is interesting beyond a game way
If you wanna make a dinosaur you gotta add shoulder blades for the neck a front legs and a pelvis for the tail and back leg
14:50 unconnected joints can cause issues for the creature.
this is cool
Give da kangaroo a tail!
Add a tail to the roo. Real kangaroos have them for balance. It won't tip over with one that looks like a real roo's tail.
I don't think you should connect the muscles to every single bone tho
much wow!
2:43
XDDDDDDDDDDDDD
The more complex the brain, the longer it takes it to learn, meaning you have to give it maybe 1000 generations to learn its full potential with about 15-20 seconds per generation. This is with the most complex brain you can make within the emulator.
Glacial Emperor it crashed for me at generation 206...
I have a problem.
I made a wagon in my evolution app on my iPad, and no matter how big the brain gets, everyone in every generation has the same problem. It sticks with a single muscle movement and doesn’t change. Basically, it starts with two muscles expanded and one contracted, then it stays that way for the rest of the generation. Help plz
5:47
a spider having a skull is a terrible thought
Wow, just. Wow
*_YES, THIS IS BIG BRAIN TIME_*
Having just one or two neurons in a layer is a bad idea. Having only one in a layer means that it takes all of the values from the previous layer and uses them to calculate a single value, which it outputs. Then all of the creature's actions will be based only on that one value instead of on the multiple values from the inputs. It basically destroys a lot of the information. Having two in a layer is better, but not by much.
The thing about having a complex brain is that it takes a long time to learn. There are many more connections it needs to figure out, so it takes longer. Having more layers makes it a good amount smarter, but puts a greater degree of seperation between the inputs and outputs. Having more neurons in a layer makes it smarter too, but also increases the possibility of introducing noise into the system if it has some good parts and some bad parts. Still, up to some limit (higher for more complex creatures), increasing the complexity of the brain increases the limit of how well the creature can perform.
the problem is you d need about 1000 gen to see if it works nor not... and not only 15
6K viewer here
6:14 THE PAIN
The brain was trying to leave the body lol
2:44 those Left 4 Dead tank sounds tho...
17:45 #ACCIDENTALGENIUS
Make a ring, next make feet along the outside of the ring, then add muscles to make the feat move, finally set it to run and watch.
Michael MAnville please spell feet correctly
BILL NYE THE SCIENCE GUY!!!!!!!!
If you load a previous run, the brain will be the same as it was when it was saved. Only new evolves will use the new network.
I feel like Anthomnia would be a good Red vs Blue character
You have to have all the brain connections at 7200 to simulate the neuron connections in the homan brain and you need about 100 billion neutrons
LAUNCH THE BOMBS!
4 then 2 then 1. It doesnt do much, but it gets more advanced. " This just in, Kangaroos have taken over the world with advanced standing!
I don’t think i can sleep tonight now
Grimlock doesn't like big science!
It’s EXACTLY like that dumb “Lucy” movie, where she uses 100% of her brain and just turns into god or whatever
Nerdling we already do that lol
His explanation of the neural network is really funny
Fibonacci that neural network. If it's good enough for nature to work maybe it will work here.
Increase the number of creatures per generations with the bigger brain. That gives it a better chance to do it right
You should connect two bones with only a muscle
Haha cool.
Has anyone ever seen a game that involved giving a green core creature limbs using bones and muscles much like this, but instead of machine learning, you then proceeded to control the creature and attempt to navigate through obstacles?
spore?
@@skimaskalx no no like, it was a 2d creature builder, I think at base the creature was just a green eye, then you could attach bones and join them up with muscles that you could expand and contract with keybinds? I think the story was that the creature was trying to find her sister or something?
Just imagine if these were real. All the bones scraping concrete with no skin and just bones and muscles 😣
Is this IGP’s second channel?
The brain tells the animal what of all the options, to choose, the more strings, the more options or opportunities to do something unique, like walk
You sound like a scientist
Scrap man built the best climber I bet
The AIs are glitching out because if they break themselves and make nodes or bones fly forward, the system says they went a long distance and, therefore, were the best and let’s those AIs effect the following generation the most.
What is the import file?
if humans where added as a guy in the game it would have hundreds of those little lines for it's brain
That kangaroo looks like a slug
Bruuuuuuuuuuuuuuiu o mi god!!!!
Maybe the layers are how many options of use it has for a muscle and the numbers with no names are which muscle its for