YTread Logo
YTread Logo

Much bigger simulation, AIs learn Phalanx

May 19, 2024
alien versus prey was an experiment I wanted to do to see if very simple rules could lead basic AIS to perform some tasks like chasing a target or dodging threads the rules were as follows there are two types of entities: predators and prey which has a prey live enough to reproduce thanks a predator has to eat the prey to reproduce the predator dies if it fails to eat the prey each entity has a neural network that is used to make decisions based on what it sees of its environment this neural network is evolved randomly when the entity creates an offspring in this first version, the entities can only see those of the other group, which means that they feed on easy predators and the predators only see prey.
much bigger simulation ais learn phalanx
The size of the map was 128 by 128 units and the antiages were one unit long. The

simulation

required some parameter adjustments to have something balanced, this was quite challenging due to the very small map size and anti-discounting. I also decided to go for a borderless world to avoid large build ups on the edge of the map after having found some pretty good setups for The

simulation

I let it run and got some interesting results at first, the entities are a bit less they mostly act strangely random shape a few minutes later. Predators begin to get better at chasing and eating prey, causing the prey population to collapse as there are fewer prey per year.
much bigger simulation ais learn phalanx

More Interesting Facts About,

much bigger simulation ais learn phalanx...

Many predators begin to starve under heavy selection. The same goes for prey, only the most killed survived when the predator population was at its peak. The consequence of this crisis is a more capable general population because all entities are descendants of strong foreign survivors. The cycle repeats until one of the two groups disappears. The simple approach of this first simulation showed that it was possible to get good results with reasonably skilled eyes from some basic rules, however it lacks the notion of good behavior because the entities cannot see their peers, which is why I decided to create a second version to implement a more complex perception system where entities can see both their prey and predators.
much bigger simulation ais learn phalanx
This new system requires using a more complex neural network for lower entities to understand what they are seeing to achieve this. I added a second entry for each tray thrown from the entity, in addition to the distance to the nearest heat, the value describing the nature of the target provided -1 if the object is dangerous plus one if the object is full zero if the object is neutral from the neural network it is more complex. I also wanted to have a larger simulation with many more agents to speed up the evolution process. You can watch my latest video if you want more details on this topic.
much bigger simulation ais learn phalanx
The new simulation is 16 times larger with a map size of 512 x 512 units and up to 16,000 features. Let's run the simulation with these new parameters and see what happens overseas. We can already see that some groups of entities are emerging, but it is not very sophisticated yet. I felt a little disappointed because nothing interesting was happening. It just seemed like it was the same stimulus but

bigger

. Here we can see some small groups wandering around, but nothing really impressive again. We can see some vague group movements with Predators forming a kind of front that is still not revolutionary in the end, this new version is not. what I would call a success.
I spent some time wondering why I couldn't observe more abusive group behavior and came up with two possible reasons: maybe I'm just unlucky and a simulation with a different seed would have produced a better result or maybe the Entities simply have no benefit from being grouped together that would prevent them from developing this new ability. I decided to go with the second option since it was pretty clear that predators had no reason to stay packed and share resources when they could eat everything. their path and their prayer didn't have

much

reason to stay together other than to provide an extra free meal.
This reasoning led me to the conclusion that to give predators and their prey a reason to stay packed, I should allow the prey to defend themselves in this way. They have to stay together if they want to maximize their chances of survival when attacking their prey, and similarly, prey would be more likely to survive predator attacks if there are several to defend themselves in the next simulation. Predators need to attack their prey twice to eat them. and pray on the other hand can neutralize a predator in four blows foreign let's see what happens with this new mechanism since in my previous attempt we can quickly observe small groups that emerge by looking at the population tables we can see that predators have more difficulty expanding in compared to the other foreign simulations the best strategy for them seems to be to surround large groups and hunt isolated prey foreign foreign thanks at this point it seems that the predators found a better technique to attack large groups of prey forming a denser front line thanks some of them They seem to be doing some synchronized attacks focusing on isolated prey Strange groups of predators are now roaming abroad It's really fascinating to see swarming mechanisms emerging from the same simple rules as before That's what I really like about evolutionary projects You can be really surprised by your own foreign creation thank you foreign foreign foreign foreign strategy here but suddenly they started decimating the sentence at a

much

higher rate at this point it's as if the prey are no longer doing any harm The predators even try to surround the preference side in a few seconds the 4000 praises are completely destroyed alien alien alien alien alien alien alien alien At this point the predators are using the scanner strategy by methodically raking the map.
I kept seeing this for several minutes and it seems to stay like this forever. It was actually more spectacular than I expected, but there are still some problems, the main one being that the entities cannot communicate with each other. I think this is a real problem for prey that have a very limited range of vision when in the middle of a large group. It's something I want to implement later when I find an efficient way to do this. thank you foreigner thank you foreigner thank you very much for watching this much longer video foreigner

If you have any copyright issue, please Contact