BIGGEST BRAIN EVER & WALKING HUMAN - Neural Network MAXED - Evolution Simulator (Evolution Gameplay)

Evolution Gameplay Simulator - Creating Complex Neural Networks to Make Human Creature Design Walk through Evolution! - Welcome back to Evolution! Today in Evolution, we take a more in-depth look into the Neural Network Settings, and do some tweaks to finally make our human design walk! Let's play Evolution!

Want more Evolution gameplay? Let me know in the comments!

----------

About Evolution:

Use joints, bones and muscles to build creatures that are only limited by your imagination. Watch how the combination of a neural network and a genetic algorithm can enable your creatures to "learn" and improve at their given tasks all on their own.

The tasks include running, jumping and climbing. Can you build the ultimate creature that is good at all of the tasks?

----------

Want more IGP? Check out these links:

Want more awesome indie games?

[keiwando]

Your explanation of everything was really good! Even if some details might not have been perfectly accurate, overall it was pretty much correct. Thanks for that!

One thing I want to add which I don't remember how clearly I pointed it out in the help text is that the initial random brains can have a high impact on how the evolution process is going to go. You've already correctly figured out that for the most part you can tell relatively early on if something interesting is going to happen during the rest of the simulation or not. I see people in all kinds of places say that "these types of networks and simulators don't work under 100 generations", which is not true at all. You can usually have a pretty good idea of where things are heading by about generation 10.

But, only because your creature doesn't really do anything interesting after 10-15 generations doesn't necessarily mean that there is a problem with your design. You might have just gotten unlucky with the initial brains. A good thing to do if you think your design is much more promising than what you are observing the creatures do, just go back and restart the simulation. That way all of the creatures will receive a new random brain and there will be a higher chance that at least some of them now do something interesting.

It's really the same as if you had an extremely large population and thus a lot more opportunity for variation, but since there are so many more possible behaviours (which include a bunch of variations of "just sit there and do nothing") than the number of creatures your computer could ever handle at once, your best bet is to restart the simulation a couple of times before you make your final call on whether your design was good or not.

Hope that made things a bit clearer.

A little info on the Neural Network: Each neuron (node) in the network receives a number of inputs from each neuron of the previous layer, weights them, sums them up and runs them through a so called activation function, which determines the output of that particular neuron. This output is then fed as an input to the next layer (or in the case of the final layer directly determines the muscle extension/contraction).

This is implemented using matrix multiplication. The matrices contain the weights for each input of each neuron. If your previous layer has 100 nodes and the current layer also has 100 nodes, that means that the weight matrix has the dimensions 100x100 (since the layers are fully connected). This means you would have 10000 weights - and that is only for two layers - which the algorithm would have to try to optimize over time.

As another sidenote, matrix multiplication is one of the things in computer science that people have been trying to optimize for a very long time now. The naive approach has a runtime of O(n^3), which in simplified terms means that if you make your matrix 10 times bigger, the matrix multiplication is going to take about 10^3 = 1000 times longer (asymptotically). There are some algorithms that are faster for really large matrices but it's very unlikely that anyone is going to find an algorithm that runs in O(n²). So that's where the performance impact of increasing the size of the network really comes in.

[traderjoestotebag]

The one with the most complicated brain just stood there thinking “What am i....”

[liam]

He protec
He attac
But most importantly
He muscle contrac

[MrPavelOK]

At some point the brain becomes so smart, it doesn't even bother running, because there no meaning to it.

[asluup]

“That big old brain the human has!”

*doesnt even have a solid skull*



Now that is why I love this channel :D

[devoidofvoltz]

I think the biggest problem with your two-legged half humans wasn't the muscles, but rather the gravity.

The triangular thing on the top simply made it too unstable and prone to falling. If you made the feet larger (and heavier) on the very bottom then the center of gravity would also be lower and grant increased stability to promote bipedal walking.

[NightClawprower]

Please tell me I'm not the only one who saw that he forgot a muscle in the legs...

[christiandinn]

The second leg on the overlayed design didn't have a muscle connecting the calve to the thigh

[crystal_wolfy]

I wish schools would use this game in classes like biology or maybe even programing

[SinglePlayerBRS]

the more complex the neural network is, the more iterations you will need to make it fit, but it is less limited.
Everytime you have more neurons you'll need to expect a longer learning curve, but a more optimal one

[jmmacd]

You forgot to add a muscle that PULLS the foot back like an ankle

[blepblops]

Let me explain this
A Layer is like a filter the more there are the harder it is fir the signals to get trough and the slower it evolves
A Node is like a calculation that then sends an output to the ones in the next layer and ecentually to a muscle

[veldrovive]

Just to give a little bit more information about the neural network and how each muscle decides what to do, here is a very brief synopsis of how an evolutionary dense neural net like this works. Let's take the example of the simple walking person at the beginning of the video. The six input nodes represent six different numerical values whether that is how many nodes are touching the ground or some sort of oscillating clock or any other variable. In a dense net, each of those input nodes connects to each of the nodes in the next layer where each connecting line indicates a weight. The input number gets multiplied by the weight in the line and then goes into every node in the next layer. Every node contains something called the activation function which basically does some sort of mathematical operation on the numbers it receives from the last layer and then that new layer becomes the input and passes along its values to the next layer via the connection weight. The same process happens at each layer until the last layer acts as the output and spits out a value for how much contraction or expansion the muscle should give. The new network is tested by running the simulation and then, since this is an evolutionary system, a new set of brains is constructed that have slightly different values from the best brains of the last generation.

[Lovinia]

Big brain boi just stood there.
He trying to fly by brain power alone

Hop man though.. he’d have a fine butt with all those squats

[dorsk]

2008: Man plays QWOP
2018: Machine plays QWOP

What a time to be alive.

[jeffreyohler]

11:56 *"Yo check out my headstand!"* He created a B-boy Breakdancer!!!

[thefluffy]

IGP, I don't see the problem. I walk like this everyday. 0.o

[tpomr]

Thincc with the thiccest of cranium muscles

[toni]

i would say layers means the amount of different joints that the creature has to move, example: in a human step the hip, knee and ankle need to move so it needs at least 3 layers to do so, in the "modak" that you use only 1 layer is required since its movement only requires one joint to move

[ourtube]

His brain is even bigger than mine..
although thats not much of an achievement lol