Artificial Life Predator Brain - Neural Net Breakdown
Вставка
- Опубліковано 26 сер 2024
- In this video we observe a carnivorous artificial lifeform in NeuraQuarium, then break down some of the structures in its neural network that help explain its behavior.
You can find "Sharks.save," the scenario from this video; a free demo of NeuraQuarium; and more information here: urocyongames.i...
For more detailed information about how NeuraQuarium's neural nets work, check out this (amusingly titled) forum thread: itch.io/t/2535...
That's a neat brain visualization with the neurons grouped by function. That and highlighting the paths make it really easy to understand what's going on in the neural network.
I've never subscribed faster. I LOVE evolution simulators and I love this project! Have you considered giving the plants their own set of narrow traits and reproduction methods so that they can evolve in and change location with the creatures. So that you don't have to change the environment by hand.
Have you ever consider to make the neuron position non static in the brain visualization? We can consider the connection between neurons as a spring. And as it has higher weight so stronger the spring is. It seems this can make better a neuron clusters visualization. For example the "hunter center" becomes easier to see since the high weighted connections attracts neurons of it
That's an interesting thought -- it would probably have to be a separate screen, since it would probably take up more space, but it might be neat!
Sounds similar to one of the methods used to visualise high-dimensional data
Amazing how such a small brain can have such complexity of behavior.
Not really, as the behaviors are probably hard-coded (although the mappings from the states to the actions are not). But a very very impressive project nevertheless.
I was wondering. Real fish usually travel in large groups and disperse when a predator approach the group. It would be cool if their brains could by accident be doing strange things like follow a kin sometimes when they are not hungry. Or in some circumstances even swim toward predators or just colors. With more opportunities for utter nonsense behavior, the more opportunities to evolve effective but counterentuitive survival strategies. I guess slower rate of mutation and larger brains are needed
All those behaviors are totally possible within the existing system. Well, not recognizing kin per se, but swimming toward colors, following scent trails (which could include leaving a musk trail for others to follow,) and so on. It's very confusing and difficult but the game does allow you to edit the neural networks as well, so theoretically one could "test out" these behaviors. But they also have the potential to evolve naturally.
The more simulations I run though, I agree, brain mutation rates need to be really low for complex behaviors to sustainably evolve.
@@urocyongames yes, im not knowledgeable about neural networks, so my comment was just guessing about what behavior this system is capable of. I trust you know it is plenty enough. And of course, slow rate of mutation is boring, but maybe that is necessary to find stable and unique species behavior
@@urocyongames do you use node mutation? Because that helps to combat the problem
I want to play this
This looks interesting. Is the brain size fixed or can that be expanded? Just worrying, that it might be to small for some more advanced or finetuned behaviour, judging from what is visible on the screen. If i remember correct, the first Creatures game used brains with 640 neurons for the concept lobe alone (ok, it was by far the largest chunk) in the Norns brains - with a total of 952 neurons, and 5000 connections. And even those ran into trouble cuz not enough space ^.^ Yeah, to be fair, the Norns had way more abilities and required way more brain power, and were usually numberwise way less in one level than what you have as population :)
The number of neurons is fixed -- as you alluded to, population sizes are the main restriction -- each brain might be small, but simulating 500 of them in real time eats up those compute cycles! The challenge, in some ways the goal, of this project has been discovering how to design the constraints and capabilities of such a small neural network such that they can still evolve interesting behaviors.
@@urocyongames Yes, i have seen that exact problem before in other projects ^.^ There was a different one, that optimized his neural network for his simulation, lets see, if UA-cam allows to share the link.
ua-cam.com/video/f_HwyDfvCZQ/v-deo.html
I want to make same thing, but in 3d, like Spore, or something
Top video, btw! Keep it up, I just saw the viewcount.
If you're looking for spore-like games then maybe check out Elysian Eclipse, Adapt or Thrive, they're all in early development but they are extremely promising, especially Thrive
@@person4579 also species alre
The biochemical smells of predators and preys is something I've never seen before. Was it inspired by some real life research paper?
The idea to have critters emit different scents depending on their diet was driven by the desire to give them a sensory stimulus that could help them determine something useful about nearby critters. I place a high priority on keeping the neural net inputs and outputs as "pure" as I can, not allowing them to "read the code" so to speak, or to have outputs that just activate some scripted behavior. Another reason I thought scent would be worth simulating is that it lingers in the water, which might allow critters to detect where something *used to be* and follow the trail. This implementation achieved both goals, as "smelling sulfur" doesn't necessarily mean there is a carnivorous critter nearby -- it just means there is sulfur in the water -- but if you are adapted to do so, you could detect a particular scent and attempt to follow it to its source.