Видео

Alternating resource locations for prey . Prey heards up predators pack up to follow them. Havent finished the video but first thing that came to mind. Why gehngis khan concord the world . Attacked fast but never slept in the same place. Hunter and prey

Great use of curriculum learning.

Sorry but i dont think they learned anything. Their behavior was just a consequence of your changes. For example: they didnt learn to form a wall and attack the prey. since they now can see each other, they cant see prey that arent directly in front of them, so the only way for them to go is for the next prey that is in front of them, and that is valid for every other predator since they also can see only a narrower angle in front of them. Do you get what i mean? If they only can see foward they will go foward. Not because the AI learned to form a wall, but your changes doomed them to a wall.

Has anyone tried going from harder to easier training?

It looks like there was an exploit way back at 18:34 Two natural next steps: 1) Make a physical version, 2) Triple pendulum.

At 19:20 you can see that it looks like the predators are doing more focused attacks on small parts of the line. Kinda like a pincer maneuver

So no Triple pendulum?

Regarding the chaos, is there are good comparison to be made to the three bodies problem?

Your video is satisfying and easy to understand for me and maybe others too I'm interested in your code, but when i clone your repo i get tens of errors in vs code Let me know if there's something to do before compiling

So does this mean one could build a machine in the real world. Plug this code in and it would be able to balance a double pendulum? :O

VTubers we need

8:03 the neural network appears like a rocket XD

what tools are you using for this type of development?

Feel like I just watched a simulation of white blood cells attacking viruses haha

Your graphics are amazing! How do you make them?

Which software are you using?

What software does this use?

Perfect video 👏

Don’t know if it’s just the colors you used, but it’s strange that 4000 pendulums look 3d

So the prey project is actually the basic template of how life began.. Simple single cells adapting to work together to become more complex organisms to stop predator cells from eating them and eventually becoming roaming predator organisms trying to consume other organisms

Make apex predators

It's super nice how changing very little to the perception leads to flocking like behaviour / simulation

Can’t you solve the same task with a traditional closed loop controller?

It'd be interesting to see what happens on a perfectly symmetrical map. I think the colonies with slightly closer food do better.

Hey there, I stumbled upon your video and I thought I might just ask you... For quiet a while I was playing with the idea of training a algorithm on solving the kinematics for my joint that is currently running as a simulation in Unity. The kinematics are quiet complex and I am struggling to define them mathematicaly so I was thinking that training an algorithm would also be a possible solution? What do you think? What would be your approach to the problem? 😊

How do you automatically create the hidden layer?

can't any entity detect its own kind??

🔥🔥

Thank you

GREY GOO?

DIDINIUM ABUSE

I have a question - its been on my mind a few years, maybe you can help. Take two such ('identical') chaotic pendula. Set them off together and measure the time it takes for their paths to diverge. The longer time taken, the smaller the difference in starting conditions. right? So couldn't we use this kind of system (Theoretically) to determine very small (microscopic) distances? Any thoughts??

I laughed at the end when you said "rudimentary" algorithm :) Anyhow - Loved this journey with you. Really fascinating. Thanks for sharing (Seems like you have over-come Loki...)

We need a 10 hour version of the double pendulum trajectory.

Can you explain better which kind of reinforcement learning agents are you using and how do they decide the network architecture? also for curiosity, what graphics visualization tool are you using ? thanks a lot man

Congrats on the video man! Literally the best 20min in RUclips since many time ago!

Two things Im curious about. 1. Is that configuration stable to small pertubations in the middle pendulum 2. Would the nn still manage to work if instead of gravity and friction, the relative length of each pendulum increased. Starting from basically a single pendulum and gradually the second pendulum increases length and the first decreases until equal in length

Really great video. I find it counterintuitive that the networks have so many layers without much interconnection between the middle layers. Is that a product of your learning process?

excelent video. I learn a lot. Do you have any video where you explain the code?

Now do it with a triple pendulum

Now 3 😂

Imagine this as a desk toy.

This is the same way humans should be trained

when will be a github repo?

Why the heck using an AI for that task? There is a much simpler solution for it. Just use an controller.

Absolutely beautiful! Impressed by the great work. Also, I'm wondering what tool or module was used to change the "nodes" model during the simulation... Anyone knows ?

lenght

which software do you use?

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

Hey, can you open source it?