First of all, install dependencies with pip install -r requirements.txt.

To launch the simulation, run main.py with python main.py. Before doing that, you might want to change options. At the top of main.py, you can modify the following code :

device = 'cuda:0' # Device on which to run the automaton
W,H = 600,600 # Size of the automaton
dt = 0.1 # Time step size


interesting_dir = os.path.join('remark_record') # Directory containing the parameters to load when pressing 'm'
remarkable_dir = os.path.join('data','remarkable') # Directory containing the parameters to save when pressing 's'

Leave interesting_dir and remarkable_dir as default for now.

Run the main.py : python main.py. A window should open, that's the Lenia world. To make actions, use the following hotkeys :

• n : Changes to random new parameters. Sampled from param_generator in batch_finder.py
• m : Load random phase transition parameters (or any saved parameters in interesting_dir)
• u : Slightly vary around the existing parameters. Can be pressed repeatedly.
• j : Re-initialize state with Perlin noise
• s : Save the current parameters inside remarkable_dir
• p : Pause/unpause simulation
• k : Display current
• r : start/stop recording video. Video will be saved in ./videos when closing main window.
• delete : set the state to the dead state.

To explore remarkable parameters we found, launch the main.py, press m and subsequently j to seed the initial condition. Behold the creatures living in your computer ! Each time you change parameters with m, do not forget to reseed the initial condition with j. Use n to explore parameters sampled randomly from the prior region defined in param_generator.

The generation of phase transition parameters using PTF is done with batch_finder.py. Note that for the generation to not be too slow, it is recommended to have a GPU for this.

The parameters than can be altered are present at the beginning of the file :

#============================== PARAMETERS ==========================================================

# Where to save the found parameters
folder_save = './data/paper_search'

device = 'cuda:0'
H,W = 200,200 # Size of the automaton
dt = 0.1 # Time step size
N_steps = 800 # Number of steps to run the automaton for

num_points = 40 # Number of points to find
refinement = 8 # Number of steps to run the dichotomy search for
cross=False # If True, will compute the transition point between all pairs of parameters. Useful for huge generations, but lessens variations
use_mean = True # If True, uses the mean of the activations to determine death. If False, uses the max.

# threshold below which we say we have found a dead config in the initial search
threshold_e = 0.05
# threshold below which we say we have found a dead config in the dichotomy search (generally matches threshold_e)
threshold_i = 0.05

batch_size = 20 # Number of worlds to simulate in parallel. Reduce if you run out of memory

# Uncomment to use the equivalent of a 'TEMP' directory. IS EMPTIED EACH TIME THE SCRIPT IS RUN
folder_save= 'data/latest'

Important parameters to note are :

• folder_save : folder location in which to save the found parameters. Note that parameters will be saved as duplicate, once in folder_save/batch as 'batched' parameters, and once in folder_save/individual as 'individual' (unbatched) parameters.
• device: set this to 'cpu' if you don't have a GPU, otherwise to the ordinal of your GPU ('cuda:0' if only one GPU)
• N_steps : number of steps to run the automaton before determining the phase.
• num_points : total number of phase transition point to generate
• batch_size : Number of world to generate in parallel. If the program crashes with a CUDA error, lower this.

Not the last line; if this is left uncommented, it will save the parameters in data/latest instead of the specified folder. Each time batch_finder.py is run, this folder is emptied. Useful to try a lot of different generations !

Finally, the function param_generator is defined below. This function defines the prior that generates the parameters. It can be adjusted if you want to change a prior. Be sure to return a dictionary containing tensor in the same sizes as presented.

All ready ! All you need to do is to run python batch_finder.py, which will run the PTF algorithm with the above specifications. To see real-time simulation of the discovered parameters, head into main.py, and change interesting_dir to the value of folder_save in batch_finder. Now, when pressing m, you will see the dynamics for the parameters you just generated !

Note : the random prior parameters are also save in 'data/latest_rand'.

To generate videos, you can either use the r key in main.py and record live simulation, or you can use the file indiv_video.py. The parameters are as follows :

param_dir = 'data/latest_rand/individual' # Directory containing the individual (unbatched) parameters
out_dir = 'test_videos' # Directory to save the videos


simulation_time = 1800 # Number of frames to simulate
size = 500,500 # Size of the simulation
fps=120 # Framerate of the video
device='cuda:0' # Device on which to simulate

Most are self-explanatory. Set param_dir to the folder containing a set of individual parameters, choose the output directory, and run the script. It should generate videos of the dynamics. It takes a while to run !