Much bigger simulation, AIs learn Phalanx

foreign versus prey was an experiment I wanted to do in order to see if very simple rules could lead basic AIS to perform some tasks like chasing a Target or dodging threads the rules were the following there are two types of entities predators and prey a prey has to live long enough to reproduce thank you a predator has to eat prey to reproduce the Predator dies if it doesn't manage to eat prey each entity has a neural network used to make decision based on what it sees of its environment this neural network is randomly evolved when entity is create an offspring in this first version entities can only see the ones from the other group meaning that prey on easy predators and predators only see prey [Music] the map size was 128 by 128 units with antities being one unit long the simulation required some tweaking of the parameters to have something balanced this was quite challenging because of the very small Map size and anti-discount I also decided to go for a borderless world to avoid big stacks on the edge of the map [Music] after having found quite good settings for the simulation I let it run and got some interesting results at first the entities are a bit less than mostly act randomly [Music] foreign [Music] a few minutes later Predators start to become better at chasing and eating prey resulting in the prey population collapsing [Music] since there are fewer prey a lot of predators start to starve applying a strong selection on them the same goes for prey only the most killed survived when Predator population was at its peak the consequence of this crisis is an overall more skilled population because all entities are descendants of strong survivors foreign this cycle repeats until one of the two groups disappears the simple approach of this first simulation showed that it was possible to obtain nice results with reasonably skilled their eyes from few basic rules however it lacks the notion of good behavior because entities cannot see their peers that's why I decided to create a second version to implement a more complex perception system where entities can see both prey and predators this new system requires to use a more complex neural network to the low entities to understand what they are seeing to achieve this I added a second input for each tray casted from the entity in addition to the distance to the closest heat the value describing the nature of the objective provided -1 if the object is dangerous plus one if the object is full zero if the object is neutral since the neural network is more complex I wanted to also have a bigger simulation with a lot more agents to speed up the evolution process you can check my last video if you want more details about this subject the new simulation is 16 times bigger with a map size of 512 by 512 units and up to 16 000 entities let's run the simulation with these new parameters and see what happens foreign [Music] we can already see that some groups of entities are emerging but it's not yet very fancy [Music] foreign [Music] [Music] I was a bit disappointed because nothing exciting was happening it just seemed that it was the same stimulation but bigger here we can see some little groups wandering around but nothing really impressive again we can see some vague group movements with Predators forming some kind of front still not revolutionary in the end this new version isn't what I would call a success I spent some time wondering why I wasn't able to observe more abuse group behavior and came up with two potential reasons maybe I'm just unlucky and a simulation with a different seed would have produced a better result or maybe the entities just don't have any benefit being grouped preventing them to evolve this new skill I decided to go for the second option as it was quite clear that Predators had no reason to stay packed and share resources when they could just eat everything on their path and pray didn't had much reason to stay together other than providing an additional free meal this reasoning led me to the conclusion that in order to give predators and prey a reason to stay packed I should allow prey to fight back this way Predators have to stay together if they want to maximize their chance of survival when attacking prey and similarly prey would be more likely to survive predator's attacks if there are several to fight back in the next simulation Predators need to attack prey two times in order to eat them and pray on the other hand can neutralize a predator in four hits foreign let's see what happens with this new mechanism [Music] [Music] as in my previous attempt we can quite quickly observe little groups emerging looking at the population charts we can see that the Predators have more difficulties to expand compared to the other simulations foreign [Music] best strategy for them seems to be to circle around big groups and Hunt isolated prey [Music] foreign [Music] foreign [Music] foreign [Music] [Music] [Music] thank you [Music] [Music] [Music] at this point it seems that Predators found a better technique to attack big packs of prey by forming a denser front line [Music] thank you [Music] some of them seem to be doing some synchronized attacks focusing isolated prey foreign [Music] [Music] [Music] foreign groups of predators are now wandering around [Music] foreign foreign it's really fascinating to see swarm mechanisms emerging from the same Simple Rules as before that's what I really like about evolutionary projects you can be truly surprised by your own creation foreign [Music] thank you foreign [Music] foreign [Music] foreign [Music] foreign [Music] strategy here but they suddenly started to decimate the prayer at a much higher rate at this point it's like prey aren't doing any damage anymore Predators even try to surround the preference side in a few seconds the 4000 praise are completely destroyed foreign [Music] foreign foreign [Music] [Music] foreign foreign [Music] [Music] foreign foreign foreign [Music] at this point predators are using the scanner strategy by methodically raking the map I continued watching this for several minutes and it seems to stay like this forever this was really more spectacular than I expected but there are still some issues the main one being that entities cannot communicate together I think this is a real issue for prey that have a very limited range of vision when they are in the middle of a big group this is something I want to implement later when I find an efficient way to do this thank you foreign [Music] thank you [Music] foreign thank you very much for watching this much longer video [Music] [Music] foreign [Music] [Music] [Music]

Channel: Pezzza's Work

Views: 2,022,668

Keywords:

Id: tVNoetVLuQg

Length: 29min 12sec (1752 seconds)

Published: Sat Nov 05 2022