An experiment with evolving creatures.
The creatures breed and try to flood you. Whoever survives before being shot or eaten is allowed to breed again.
Based on Kataster.
Inspired by the unmatched freeware version of Crimsonland.
If you like it, let me know! I'm dcz on freenode IRC, and my email is bmatic (dot) dcz (at) porcupinefactory.org .
Get assets from https://porcupinefactory.org/data/assets.zip and unpack them into the assets directory.
Then you can start the breeding process:
cargo run --release
Both the shooters and the fleas will start breeding. Fleas are pre-seeded, so they will not get much smarter.
But the shooters… They start out uncoordinated. As they mutate, and as fleas take away the dumbest, only the high scoring one will remain in the gene pool.
After about 100 attempts, the gene pool will get honed, and you should be seeing good shooters regularly.
Video of a shooter after 6 lucky mutations
Brains of all creatures are neural networks, and the connections between neurons, and neurons' activation funcions are what evolves.
The bottom left corner is the live view of the brain of the shooter.
The brain takes in signals from the two topmost circles: angle to nearest baddie, time alive. The signal passes along connections from top to bottom to neurons, and the final circles-neurons at the bottom result in the output signals: angle of the weapon, body turn rate, and movement speed.
Signal strength is expressed with color: gray means relaxed (0), yellow active (positive), blue negative. The stronger the signal is, the stronger the color.
Some connections have no circle at the top: those are "bias" connections. The source strength is always 1.
Every neuron's task is to sum up incoming signals, and then to activate based on the result. Most neurons activate proportionally, but there is also the sigmoid activation (result never exceeds [0, 1]), the step (anything below 0 turns into 0, anything above activates to 1), ReLU (anything below 0 is shunted to 0), and Gaussian (the farther the sum from 0, the farther the activation from 1).
So don't be surprised that a neuron with no connections is suddenly always active!
Yes. Let me know you care and I may fix it.
Each round, some text will print on the console, informing about the current evolution. It looks like this:
Spawn offspring of 87
Spawned genotype Mut 20
Hidden
Linear: -0.953 0.001 0.001
Linear: 0.000 -0.054 0.011
Gaussian: 0.000 0.070 0.687
Out
Linear: -0.853 0.000 -0.361 0.000
Linear: 0.000 -0.720 0.000 0.000
Step01: 0.000 1.190 0.000 0.000
Wrote shooter.dot
Preserved as 90 with score 720
Pop 22
Let's break it down.
Spawn offspring of 87
Each shooter in the gene pool gets a unique number. This shows the number(s) of the parent(s) of the just-spawned shooter.
Spawned genotype Mut 20
The shooter went through 20 mutations. Mutations will happen more often when population is low.
Hidden
Linear: -0.953 0.001 0.001
Linear: 0.000 -0.054 0.011
Gaussian: 0.000 0.070 0.687
The hidden layer of neurons in the brain (rows). Inputs are: angle to baddie, time alive, bias (columns).
Out
Linear: -0.853 0.000 -0.361 0.000
Linear: 0.000 -0.720 0.000 0.000
Step01: 0.000 1.190 0.000 0.000
The output layer. Inputs come from the previous layer of neurons, plus an extra bias comes last (columns). Outputs are rows: weapon angle relative to movement direction, body turn speed, walk speed.
Wrote shooter.dot
The shooter was saved as a file for graphviz. Convert it to png using:
dot shooter.dot -Tpng -oshooter.png
Next comes the notice about shooter's death.
Preserved as 90 with score 720
Pop 22
The shooter's ID in the gene pool is 90 in this case, resulting in 22 different genotypes in the pool.
The entire code may be distributed under AGPL 3.0 or higher, at your leisure. License text in the agpl-3.0.md file.
In addition, remaining Kataster pieces are licensed under MIT. License text in the MIT.txt file.