The hybrid between ML and NEAT is so interesting! This is something many similar videos don't talk about, I'm glad finally someone did! This is something I'm very excited to see worked on and improved. Great video!
can't wait for the evolving bodies! also what about having different environments that the creatures could live in, like a grass land or a lake? I don't know how doable it is, but I think it would be interesting to se how different optimal strategies exist for different environments.
I have an idea planned for that and it isnt really like having different “biomes” but more like the laws of physics are different in certain areas. So for example in one area moving costs less energy, the “water” they swim in is more viscous, they think slower, vision is distorted or even time itself is faster.
@@scasz what I'm curious about is how the neural networks would learn to avoid leaving the area if it is implemented or how they will know what area there in, because from my experience with that type of data being fed into neural networks, they often struggle to decode it and it becomes noise. I guess there is also the route of having the area be big enough so that even if creatures leave and die, there's plenty left inside the area still.
@@_ERR_404_ Im probably go for the creatures being able to "feel" which areas they are in so they will have an input for stuff like viscosity and time multiplier but the areas will be large enough so there will always be some creatures as I want to see if I can almost create two environments with their own ecosystem that has an invisible barrier separating them that some creatures do breach and maybe survive to reproduce, having them too small wouldn't really give the creatures space to thrive without competition. The new hybrid brain should hopefully be able to learn which areas are good and bad but if not then the areas are big enough. If I optimise everything good enough I should be able to make the simulation much larger
@@scasz I wonder if the creatures would just evolve using their position to recognize if they are too far away from the center of the area and then turn back if they are. I mean if that's the case I wonder if the ai would then need to be changed at all.
That could be possible and it could almost recreate how turtles go back to where they were born but just because an environment is different that doesnt make it bad as the creature might even thrive more in it than their original one.
I'm really looking forward to your next video! I was huge into Bibites and I would love to see another project like it. One of the biggest things that would make a difference is evolvable neural building blocks. The same blocks could be used multiple times across the brain--when used it would take inputs run it through and output it. The blocks themselves could run "fresh" each time--as if they were a template.
Thats pretty interesting, i actually had a similar idea but didnt know if it already had a name so I went with a more tradition RL network as I saw more info on it. Will definitely check it out
@@scasz you accidentally reinvented Hebbian learning on your own? 🤯 It seemed to work well on Creatures norns. I wonder if you can also compare Hebbian learning all the time vs only Hebbian reinforcing when good stuff (mating feeding etc) and negative reinforcement when bad stuff (taking damage etc). Given how simple the creature tasks are, it makes sense that NEAT would be more efficient at solving it than reinforcement learning/gradient ascent. It's basically a linear control task (turn left if food or mate is at left and vice versa). NEAT is still super efficient at tricky stuff such as swing-up double inverted pendulum, so perhaps PPO would shine if the environment or movement are trickier, such as self-balancing increasing value of unchosen resources, undulating segmented swimming, etc
The planned evolving bodies will need more powerful brains that can learn faster so thats why im doing all this (maybe hebbian could be a better alternative to traditional rl) as I want to have bodies with individual segments that can be moved to swim, bite, etc
@@scasz good idea, learning during the creatures' lifetime instead of just through evolution could speed up the evolution by a lot via Baldwin effect. Still, variable topology genetic algorithm seems to be super efficient at finding the optimal topology, so I wonder how NEAT genetic topology + Hebbian learning would compare to Norn's learned topology + Hebbia learning
@@scaszdid you use node mutation (only mutate the incoming weights of some nodes at a time), or all nodes at once? The former is safer and faster, with less risk of bad mutations cancelling out the good ones
The hybrid between ML and NEAT is so interesting! This is something many similar videos don't talk about, I'm glad finally someone did! This is something I'm very excited to see worked on and improved. Great video!
This channel is so underrated
Did you played bibites: digital life?
WOO hes BACK
Nice Video man. ❤
can't wait for the evolving bodies! also what about having different environments that the creatures could live in, like a grass land or a lake? I don't know how doable it is, but I think it would be interesting to se how different optimal strategies exist for different environments.
I have an idea planned for that and it isnt really like having different “biomes” but more like the laws of physics are different in certain areas. So for example in one area moving costs less energy, the “water” they swim in is more viscous, they think slower, vision is distorted or even time itself is faster.
@@scasz what I'm curious about is how the neural networks would learn to avoid leaving the area if it is implemented or how they will know what area there in, because from my experience with that type of data being fed into neural networks, they often struggle to decode it and it becomes noise. I guess there is also the route of having the area be big enough so that even if creatures leave and die, there's plenty left inside the area still.
@@_ERR_404_ Im probably go for the creatures being able to "feel" which areas they are in so they will have an input for stuff like viscosity and time multiplier but the areas will be large enough so there will always be some creatures as I want to see if I can almost create two environments with their own ecosystem that has an invisible barrier separating them that some creatures do breach and maybe survive to reproduce, having them too small wouldn't really give the creatures space to thrive without competition. The new hybrid brain should hopefully be able to learn which areas are good and bad but if not then the areas are big enough. If I optimise everything good enough I should be able to make the simulation much larger
@@scasz I wonder if the creatures would just evolve using their position to recognize if they are too far away from the center of the area and then turn back if they are. I mean if that's the case I wonder if the ai would then need to be changed at all.
That could be possible and it could almost recreate how turtles go back to where they were born but just because an environment is different that doesnt make it bad as the creature might even thrive more in it than their original one.
What's the music you were using for the timelapses in the video?
I feel like I remember listening to it somwhere.
Ive updated the description to include the name and links to the music I used. It was probably either skystrike or sharp edges
Such a cool channel
Amazing Video
I'm really looking forward to your next video! I was huge into Bibites and I would love to see another project like it. One of the biggest things that would make a difference is evolvable neural building blocks. The same blocks could be used multiple times across the brain--when used it would take inputs run it through and output it. The blocks themselves could run "fresh" each time--as if they were a template.
Neat video 👍
All hail the algorithm gods
TUTORIALS PLS
What about, instead of PPO, you use Hebbian learning on NEAT networks?
Thats pretty interesting, i actually had a similar idea but didnt know if it already had a name so I went with a more tradition RL network as I saw more info on it. Will definitely check it out
@@scasz you accidentally reinvented Hebbian learning on your own? 🤯
It seemed to work well on Creatures norns. I wonder if you can also compare Hebbian learning all the time vs only Hebbian reinforcing when good stuff (mating feeding etc) and negative reinforcement when bad stuff (taking damage etc).
Given how simple the creature tasks are, it makes sense that NEAT would be more efficient at solving it than reinforcement learning/gradient ascent. It's basically a linear control task (turn left if food or mate is at left and vice versa). NEAT is still super efficient at tricky stuff such as swing-up double inverted pendulum, so perhaps PPO would shine if the environment or movement are trickier, such as self-balancing increasing value of unchosen resources, undulating segmented swimming, etc
The planned evolving bodies will need more powerful brains that can learn faster so thats why im doing all this (maybe hebbian could be a better alternative to traditional rl) as I want to have bodies with individual segments that can be moved to swim, bite, etc
@@scasz good idea, learning during the creatures' lifetime instead of just through evolution could speed up the evolution by a lot via Baldwin effect.
Still, variable topology genetic algorithm seems to be super efficient at finding the optimal topology, so I wonder how NEAT genetic topology + Hebbian learning would compare to Norn's learned topology + Hebbia learning
@@scaszdid you use node mutation (only mutate the incoming weights of some nodes at a time), or all nodes at once? The former is safer and faster, with less risk of bad mutations cancelling out the good ones
Nice
A good
5th
Third