Critters must catch food (green circles) and avoid dangers (red squares). The brain of each critter is a small feedforward neural network. Brains are trained using a genetic algorithm to simulate evolution. Every 20 seconds, the five best critters of the latest generation are sent to the simulation running in the graphical user interface (GUI).
This software is intended to be built on a Unix-like operating system with GCC. It is known to build and run on Linux and MacOS X.
This software has a dependency on the SDL library for the graphical user interface (GUI), so please ensure the SDL library is installed before building. Some Linux distributions split run-time and developement files in separate packages. If it is the case for your distribution, you will need both. On CentOS, this translates to the following yum install command:
yum install SDL SDL-devel
An x86 processor is required because some parts of the software use compiler intrinsics for SSE2 instructions.
If starting from a clean clone of the Git repository, run the following:
./autogen.sh
Then, configure the sources:
./configure
To compile (or re-build after having made some changes to source files):
make
To install locally and then run:
make install
critters
Alternatively, you can run from the compiled source directory without installing:
src/critters
You can change the number of neurons in the hidden layer and their activation function (sigmoid-like, gaussian-like or ReLU) by changing the constants at the top of src/genome.h (and then re-building). You can also modify the parameters of the genetic algorithm at the top of src/breeder.h.
Critters have two senses: vision and smell src/scene.c. Vision is limited to a +/-63 degrees angle. Critters know the distance and angle of the nearest food item, danger and outer wall within that angle and subject to a distance limit. The closest item of each type within the visual field might not be the closest overall. Critters' sense of smell allows them to know the distance of the closest food item and danger omnidirectionally up to a distance limit (they cannot smell walls). That makes eight inputs src/stimuli.h:
- (Vision) Distance and angle of nearest food item within the visual field.
- (Vision) Distance and angle of nearest danger within the visual field.
- (Vision) Distance and angle of nearest outer wall within the visual field.
- (Smell) Distance of nearest food item.
- (Smell) Distance of nearest danger.
Each critter's brain is a small feedforward artificial neural network with a single hidden layer src/brain.c. The two outputs of the neural network are the speed of the left side and the right side of the critter. If these speeds differ, the critter turns. The weights of the neural network are part of a critter's genome src/genome.h. Also part of the genome is the colour of the critter's head. This provides a visual indication of what happens to a characteristic that is not selected for.
Training is performed with a genetic algorithm. At each generation, each critter is simulated for a short time and a fitness function is computed that depends on the number of times the critter catches food or is caught by a danger. For simplicity of implementation, critters are not eliminated outright when they get caught by a danger, but the associated fitness cost penalty coupled with the selection procedure are such that selection becomes very unlikely when that happens.
The selection procedure works as follow src/breeder.c: First, the genomes with the lowest fitness score are discarded. Then, a pool of genomes is created by picking the genomes with top fitness score, then picking a few other genomes randomly, and finally adding a few randomly-generated novel genomes. Genomes with top fitness score are added multiple times to increase the probability they get chosen. Once the pool has been created, pairs of genomes are selected randomly with uniform distribution within that pool.
New genomes are created from each selected pair by recombination with mutation src/genome.c.
Although written in C, the code is structured similarly to object-oriented software. Most header files in the source directory (src) have a similar format where a structure that represents an object type is declared first, followed by functions that act on this type. The implementation of these functions is located in the source file with the same root name as the header file. Inheritance is implemented by the structure that represents a derived object type having the structure of the parent type as its first member.
- A scene (
scene_t- src/scene.h src/scene.c):- Contains a collection of things (
thing_t- src/thing.h src/thing.c)- Things can be critters (
critter_t- src/critter.h src/critter.c)- Contains a genome (
genome_t- src/genome.h src/genome.c) - Contains the computed output of the critter's brain (
brain_control_t- src/brain.h src/brain.c)
- Contains a genome (
- Things can be objects that bounce on the scene outer walls (
boing_t- src/boing.h src/boing.c)- Things that bounce can be food (
food_t- src/food.h src/food.c) - Things that bounce can be dangers (
danger_t- src/danger.h src/danger.c)
- Things that bounce can be food (
- Things can be critters (
- Contains a collection of things (
- A GUI window (
window_t- src/window.h src/window.c):- Contains a scene that is rendered on screen.
- A breeder (
breeder_t- src/breeder.h src/breeder.c):- Contains a collection of one or more worker threads (
thread_state_t- src/breeder.c)- Contains a scene where critters are simulated
- Contains a list of critters that needs to be simulated to evaluate the fitness function.
- Contains a list of critters that have already been simulated.
- Contains a collection of one or more worker threads (
The main() function located in src/critters.c (not to be
confused with critter.c) instanciates one window and one breeder. Once started,
the breeder performs its work on one or more worker threads while the main()
function calls the update function of the window's scene in a loop. Every 20
seconds, the main() picks the five best critters from the breeder's latest
generation and replaces the critters in the window's scene with them.
I reused the AVL tree implementation src/tree.c which I wrote for another personal project that I started but never published, which is why it looks somewhat overkill for this purpose and less well integrated than the rest of the code.