Відео

This is extremely cool, yes... but who designed your UI??

Hey, can you open source it?

maybe it is a stupid question, but how can some learn such stuff, I mean I am doing my master in mathematics and computational engineering, and I have absolutely no idea of how to do this stuff. Like yeah OK , I might now have mayor problems with the maths and stuff but still, everything is so cool, beautiful and fun, I would love to learn to do such things on my free time instead solving differential equations :|

I feel like you now should write the scientific paper on this. Going from "This problem is yet to be solved" to "I solved it with about 2 hours of training for my model" is huge progress!

Loved the video, the visualizations, the breakdown and the thought of using not python for everything is whats impressive, immediate sub, hope ill learn how to do this stuff

It'd be interesting to see each predator/prey have a distinct color coded by a gene that could randomly change slightly per reproduction. That would offer more insight into the different genetic lineages.

I think if you wanted your prey to behave more dynamically, you shouldn't assume that they have an ambient food source. just creating patches of regenerating plant life that they have to migrate to would probably help.

how to download??

I've been obsessed with double pendulums for a long time now, I absolutely loved this video. I honestly started the video thinking it would be impossible to balance.

Albert the G is back

essentially: Nah, I'd Adapt

bro why there so much nerds here i just like to see a war simulation lol

Now do a triple pendulum

It would be interesting to see what other types of behaviors would emerge if the entities could destinguish how similar/different they are from each other. (Perhaps also each others age/hunger level.) Would the prey prefer to protect their closer relatives while abandoning others? Would the preditors form packs of similar kinds or would they team up with others, having different 'roles'? And/or they could have a value that they could freely turn on/off (0 or 1) or set, being able to function as a form of signal to others (and then perhaps being understood by some of the other force).

Theyre moving in herds. They _do_ move in herds...

This would be cool to put in a graph for the populations also really cool outro

The Family Guy chicken!

Would be cool to add plants that the prey can eat too.

You talk about directed graph but you have showed simple (undirected) graph. Directed graph must have arrows on eges.

I'd love to run this on my own PC for a few days and just watch it.

4:40 i hate that you end it right after it got interesting

This problem was first solved in the 1960s with adaptive filters - the real father of neural networks! 50+ years later...

pretty cool!

how big do you make the grid? 10x10, 100x100, 500x500...

the visualizations are epic

Seems like predator birthrate is too quick. Needs to be like 1/4 or lower of the prey's rate. Some obstacles, like psuedo-topography would help the prey as well

I am thinking about instead of increasing gravity and reducing friction, how about gradually increase the length of the second pendulum linkage to gradually increase the difficulty level of the training system?

My cat was very interested in the chain!

A cool idea would be to add terrain so we could see if the AI can develop guerilla/ambush tactics. Like rivers, which are slow to cross so you can wait on the other side for the prey, bolders that must be avoided, Tall grass to hide in so that the predators must really hunt for them, berry fields in which the prey multiplies twice as fast so eventually the predator can spawn camp, high ground from which you can see hiding bots and the entire map etc.

nice!

I love the interface, great video!!

Build it in real life

j'adore l'accent fr XD : )

why are u using evolutionary neurel networks? why dont u make a proper one with backpropagation?

I'm curious to know what adding a blocking mechanic might do especially if it has a specific quark like blocks all damage but only directly in front of the defender.

are anyone know what makes this amazing ui？Is it web or some lib or framework? It just amazing

wouldn't be an ai without it breaking the physics engine. Which is something ai seems to be quite reliably good at.

Thanks!

you should add an input that tells AI how much its oscillating given a time-frame. also during training you should give random "wind" inputs to the pandelum, that will teach it to perform when you "touch" it. also also, you should disable the ai for a few seconds and let it pull itself together another time, that will eliminate flukes.

Next you gotta do a triple pendulum.

Have you considered having a go at solving the three body problem?

Will there be a third part?

boooring

excuse me what

beautiful vid! keep up the good work. subbed

3h version of the thing you developp for your thumbnail please hahaha

taming double pendulam is like predicting the output of sha hashing algorithm slight variation in input data can lead to very different results each time

I'm impatiently waiting for your sequel about triple pendulum. Or maybe first you teach the double pendulum to reach a stable downward position from where it can restart.

It is adorable to see someone riding the AI wave backtrack half way into discovering model based control. But this is a lot of work though, and done really well. Here is a little input though. You basically have a physical system that can be represented in state space. For example as the angles, angular velocities, cart speed and so on. Lets say it is 5 dimensional state space. And because this is a simulation you have direct access to all the state space variables. So this is more or less what you have as inputs to your controller already. What you then want to do is to find the equilibrium you are looking for in this state space. I.e. all angles are zero, all angular velocities are zero, the cart is centered, the cart speed is zero. You then construct your objective function to be something like \sum_{t = 0}^T w_1*angle_1[t]^2 + w_2*angle_2[t]^2 + w_3*cartSpeed[t]^2 where w_x is the how much weight you put on each state variable. Now train the network. It will find something good. Now if you linearize you neural network at the equilibrium point you have a linear controller that will work as well as the network, but only when the pendulum is already stabilized. And voila, you have rediscovered Linear Quadratic Regualtors (LQRs). Linear in that it is a linear controller, quadratic in that it is a quadratic cost function you are optimizing. This is the normal way of stabilizing a double pendulum and quite straight forward (once you know the theory i guess). Now, your network is not linear, and that makes it more powerful for swinging the penduli up, which a linear controller can't do. When doing this with a linear controller, normally you have some swing up routine that brings it withing controlability of the of the linear controller, and then the linear controller takes over. I am not saying use LQR. It has already been done. I am saying take inspiration from how the objective function in LQR control is specified. It should be reusable. And then know that once stabilized, your controller will be equivalent to LQR control. Oh, one more thing. Make sure that your activation functions in your neural network are smooth. Otherwise you may never get rid of the jittery behavior.

Maybe with a Liquid Neural Network, like the Spiking Neural Network, would allow it to adapt to autocorrecting after the double pendulum is imbalanced. P.S. This would be a new tutorial of how to do a liquid neural network for UA-cam I can't find any. I appreciate you!