Master's Thesis: AI Generating Algorithms with Self-Organizing Neural Cellular Automata

The Project

This repo is code supplement for MSc. thesis at HiØ under supervision of prof Stefano Nichele for the June, 2024.

This project implements the Neural CA Frameowork where self-replication with mutation is the only way for cells to live longer. We study and Analyse the growht in Phenotypic Diversity (PD) and Genetic Diversity (GD) over longer time steps. Specifically, we intend to study, formalise and propose three tools to analyse GD and four tools for PD. Specifically, GD level tools are (1) Random Weight Selection Plot (RWSP), (2) Genotypic Hash Coloring (GHC) and (3) Combined count plot of unique colors. PD level tools include (1) Cellular Type Frequency Plot (CTFP), (2) Global Entropy Plot (GEP), (3) Gross Cell Variance Plot (GCVP), and (4) Cell Local Organisation Global Variance (CLOGV).

Quick Demo Configuration:

WIDTH, HEIGHT = 300,300
INIT_PROBABILITY = 0.1
NUM_LAYERS = 2
ALPHA = 0.5 
INHERTIANCE_PROBABILITY  = 0.2
parameter_perturbation_probability = 0.2
NUM_STEPS = 300
activation sigmoid on board
FPS = 10
budget = 4  # Cells die after 4 generations

Visit youtube channel for Demos - https://www.youtube.com/@growingnca/playlists

Quick Demo NCA Video (Redirects YouTube)

https://www.youtube.com/watch?v=kD6vfKnRA-I&list=PL_IJ0j36aBCbmN7_OvWWdjNVw-QhXFfZo&index=1&pp=gAQBiAQB8AUB

CTFP

GEP+GCVP+CLOGV - Phenotypic Diversity

Quick Demo RWSP Video (Redirects YouTube)

https://www.youtube.com/watch?v=2fDUCPR4kk8&list=PL_IJ0j36aBCbmN7_OvWWdjNVw-QhXFfZo&index=2&pp=gAQBiAQB8AUB

Quick Demo GHC Video (Redirects YouTube)

https://www.youtube.com/watch?v=KbvLTyEpMYM&list=PL_IJ0j36aBCbmN7_OvWWdjNVw-QhXFfZo&index=3&pp=gAQBiAQB8AUB

RWSP and GHC Count Plot

Research Questions:

1. What are the key species that emerge, survive, reproduce and becomes extinct (dead) in the evolved NCA model?
2. How does the genotypic diversity evolve over successive generations of the NCA model?
3. What is the impact of the self-replication and self-maintenance processes, characteristic of autopoiesis, on the genetic makeup of the evolving NCAs?
4. How do phenotypic variations manifest within the NCA model, and what are the underlying factors contributing to this diversity?
5. To what extent do the cellular activities within the NCA exhibit sensitivity to initial conditions (for example number of time steps) and how does this influence the genetic and phenotypic outcomes?

Objectives:

1. Analyze the species composition within the evolved NCA model using Cellular Type Frequency Plot and identify dominant and rare species.
2. Implement Genotypic Hash Coloring and Random Weight Selection Plot to quantify and visualize changes in genotypic diversity across multiple generations of NCAs.
3. Apply Clustering Neural Weights Approach to understand how self-replication and autopoiesis contribute to the organization and maintenance of genetic traits within the NCAs.
4. Utilize Phenotypic Diversity tools, including Global Entropy Plot and Gross Cell Variance Plot, to assess and quantify the diversity in emergent phenotypes over the course of NCA evolution.
5. Investigate the impact of dynamic landscapes on the evolution of smart and interesting phenomena, exploring adaptability and evolvability within the NCA model.
6. Evaluate the effectiveness of the proposed tools in providing fine-grained and coarse- grained analysis of Non-Uniform Self-Replicating NCAs.
7. Assess the potential applications of the research findings in advancing the understanding of life, biology, and complex systems, as well as their implications for artificial general intelligence.

Please refer to report document for further conceptual and implementation details.

Contents

• Open in Colab
• Using SLURM for Larger Experiments
• Configuration on Simula eX3
• SageMaker instance
• Experiments
• Cite this
• End Deliverables

Open in Colab

Use the colab button above to directly start reproducing the code. You can download the colab and import into your notebook / ipython environment. The advantage of Google Colab is to debug and quickly test the actual framework. We set the following parameters in first cell of the notebook and rest of the cooking material evolves the NCA and generate corresponding results.

Experiment 1

precision = 1
torch.set_printoptions(precision=precision)
WIDTH, HEIGHT = 30,30
grid_size = (WIDTH, HEIGHT)
print("Width and Height used are {} and {}".format(WIDTH, HEIGHT))
INIT_PROBABILITY = 0.1
min_pixels = max(0, int(WIDTH * HEIGHT * INIT_PROBABILITY))
NUM_LAYERS = 2 # rest hidden and one alpha
ALPHA = 0.5 # To make other cells active (we dont go with other values below 0.6 to avoid dead cells and premature livelihood)
INHERTIANCE_PROBABILITY  = 0.2 # probability that neighboring cells will inherit by perturbation.
parameter_perturbation_probability = 0.2
print("Numbers of layers used are {}".format(NUM_LAYERS))
print("1 for alpha layer and rest {} for hidden".format(NUM_LAYERS-1))
NUM_STEPS = 90
num_steps = NUM_STEPS
at_which_step_random_death = 9999999999 # Set this to infinity or high value if you never want to enter catastrophic deletion (random death happens at this generation)
probability_death = 0.004 # 40 pixels die every generation
print("Numbers of Time Steps are {}".format(NUM_STEPS))
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {DEVICE}")
activation = 'sigmoid' # ['sigmoid','tanh','noact']
frequency_dicts = []
FPS = 10 # Speed of display for animation of NCA and plots
marker_size = 1 # for plots
everystep_weights = [] # Stores weigths of the NNs from every time step.
enable_annotations_on_nca = True
budget_per_cell = 3
fixed_value = 0
budget_counter_grid = np.zeros((WIDTH, HEIGHT)) + fixed_value

With this configuration it simulates the proposed NCA framework. And then results a simulation like:

Beyond that it produces more results that are related to Phenotypic Diversity (PD) and Genotypic Diversity. PD tools are shown below:

• CTFP Plot - Frequency Count Plot

• GEP+GCVP+CLOGV (Phenotypic Diversity and Randomness)

• GD Unique Color Count Plots (GHC and RWSP)

• RWSP Animation

• GHC Animation

Experiment 2

precision = 1
torch.set_printoptions(precision=precision)
WIDTH, HEIGHT = 30,30
grid_size = (WIDTH, HEIGHT)
print("Width and Height used are {} and {}".format(WIDTH, HEIGHT))
INIT_PROBABILITY = 0.09
min_pixels = max(0, int(WIDTH * HEIGHT * INIT_PROBABILITY))
NUM_LAYERS = 2 # rest hidden and one alpha
ALPHA = 0.5 # To make other cells active (we dont go with other values below 0.6 to avoid dead cells and premature livelihood)
INHERTIANCE_PROBABILITY  = 0.1 # probability that neighboring cells will inherit by perturbation.
parameter_perturbation_probability = 0.02
print("Numbers of layers used are {}".format(NUM_LAYERS))
print("1 for alpha layer and rest {} for hidden".format(NUM_LAYERS-1))
NUM_STEPS = 90
num_steps = NUM_STEPS
at_which_step_random_death = 9999999999 # Set this to infinity or high value if you never want to enter catastrophic deletion (random death happens at this generation)
probability_death = 0.004 # 40 pixels die every generation
print("Numbers of Time Steps are {}".format(NUM_STEPS))
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {DEVICE}")
activation = 'sigmoid' # ['sigmoid','tanh','noact']
frequency_dicts = []
FPS = 10 # Speed of display for animation of NCA and plots
marker_size = 1 # for plots
everystep_weights = [] # Stores weigths of the NNs from every time step.
enable_annotations_on_nca = True
budget_per_cell = 999999999
fixed_value = 0
budget_counter_grid = np.zeros((WIDTH, HEIGHT)) + fixed_value

With this configuration it simulates the proposed NCA framework. And then results a simulation like:

Beyond that it produces more results that are related to Phenotypic Diversity (PD) and Genotypic Diversity. PD tools are shown below:

• CTFP Plot - Frequency Count Plot

• GEP+GCVP+CLOGV (Phenotypic Diversity and Randomness)

• GD Unique Color Count Plots (GHC and RWSP)

• RWSP Animation

• GHC Animation

Using SLURM for Larger Experiments

We use Simula eX3 cluster in support of Research Council Norway. We use a100q, dgx2q and hgx2q machines to run our larger experiments. The batch script looks like:

#!/bin/bash
#SBATCH -p a100q ## hgx2q, a100q,  dgx2q (old), genomaxq, xeonmazq
#SBATCH --output=run.out    ## Output file
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --gres=gpu:1
#SBATCH -t 13-00:00 # time (D-HH:MM)
#SBATCH --mail-user=XXXXX
## module load cuda11.0/toolkit/11.0.3 
## module load cudnn8.0-cuda11.0
#module load cuda11.3/toolkit/11.3.0 
#module load cudnn8.0-cuda11.3
module use  /cm/shared/ex3-modules/0.4.1/modulefiles
##module load slurm/20.02.7
##module load python-3.9.15
# ACTIVATE ANACONDA
eval "$(conda shell.bash hook)"
conda activate new_env
srun python -u run.py 

However the run.py file contains similar code that we have seen in Google Colab, but only produces large archives for evolved weights and the grid, which can further be analysed using the SageMaker notebook. For instance, the run.py file has following ingredient:

precision = 1
WIDTH, HEIGHT = 100,100
grid_size = (WIDTH, HEIGHT)
INIT_PROBABILITY = 0.0030
min_pixels = max(0, int(WIDTH * HEIGHT * INIT_PROBABILITY))
NUM_LAYERS = 2 # One hidden and one alpha
ALPHA = 0.6 # To make other cells active
INHERTIANCE_PROBABILITY  = 0.2 # probability that neighboring cells will inherit by perturbation.
parameter_perturbation_probability = 0.2
NUM_STEPS = 1000, 2000, 3000, ..., 10000
num_steps = NUM_STEPS
activation = 'sigmoid' # ['relu','sigmoid','tanh','leakyrelu']
FPS = 1 # Speed of display for animation of NCA and plots
marker_size = 2 # for plots
frequency_dicts = []
everystep_weights = [] 
ca_grids_for_later_analysis_layer0 = []
ca_grids_for_later_analysis_layer1 = []
...
# Evolve NCA weights and grid
...
with open('frequency_dicts.pkl', 'wb') as f1, open('everystep_weights.pkl', 'wb') as f2, open('ca_grids_layer0.pkl', 'wb') as f3, open('ca_grids_layer1.pkl', 'wb') as f4:
        pickle.dump(frequency_dicts, f1)
        pickle.dump(everystep_weights, f2)
        pickle.dump(ca_grids_for_later_analysis_layer0, f3)
        pickle.dump(ca_grids_for_later_analysis_layer1, f4)

We release all required code in the main runs folder. The requirements to start with the SLURM code is as follows.

• Create environment and install using requirements.txt

conda create --name myenv 
conda activate myenv 
conda install --file requirements.txt

• requirements.txt

# NumPy for numerical operations 
numpy==1.21.0 
# Matplotlib for plotting 
matplotlib==3.4.3 
# PyTorch for machine learning 
torch==1.10.0 
# Pandas for data manipulation and analysis 
pandas==1.3.3

Configuration on Simula eX3

Following configuration was used with multiple nodes of each platform to fasten the multi-fold experiments.

Platform	Model	GPU	Processor
x64_64/A100	hgx2q	g002	DualProcessor AMD EPYC Milan 7763 64-core w/ 8 qty Nvidia Volta A100/80GB
x64_64/V100	dgx2q	g001	DualProcessor Intel Xeon Scalable Platinum 8176 w/ 16 qty Nvidia Volta V100
x86_64/cpu	xeonmaxq	n022	DualProcessor Intel Xeon Max Q with mix configuration
x86_64/cpu	genoaxq	n021	DualProcessor (Epyc or Xeon) with mix configuration

SageMaker instance

We use AWS SageMaker for the easiness to process experimental results. Specifically, for the exports that we receive from SLURM experiments, which are as pickle files containing evolved weights and stored grids. SageMaker notebook instance used with almost 500 GB of storage and following system configurations as per the size of experiment exports.

Instance Type	vCPU	Memory	Price per Hour	Experiment size
ml.r5.large	2	16 GiB	$0.151	<500 MB
ml.r5.xlarge	4	32 GiB	$0.302	<1 GB
ml.r5.2xlarge	8	64 GiB	$0.605	<2 GB
ml.r5.4xlarge	16	128 GiB	$1.21	<10 GB
ml.r5.8xlarge	32	256 GiB	$2.419	<15 GB
ml.r5.12xlarge	48	384 GiB	$3.629	<20 GB
ml.r5.16xlarge	64	512 GiB	$4.838	<30 GB
ml.r5.24xlarge	96	768 GiB	$7.258	<100 GB

This classification of Experiment Size is done to avoid runtime collapse because of the memory outage during the processing of the results for the experiments. We could have produced results on runtime while experimenting using SLURM experiments, however to modularise and balance the workload, we preferred to save the weights for one time, and then process results later if required. This also brought flexibility to improve existing code bases for processing results and add more tools later for our project. We release 2_sagemaker.ipynb to process the experiments. This notebook takes public URLs of the experimentation and process results. We release these large experiments as public urls below. These publics links are also self-contained in order to check results of those expierments. You can download sagemaker notebook from this link - Download Notebook

Experiments

Before we go ahead, Full forms of corresponding short forms used in other tables:

Short Form	Full Form
$exp$	Experiment Number
$W$	Width of NCA
$H$	Height of NCA
$init$	Initial seeded agents percentage
$ppp$	Parameter Perturbation Probability
$ip$	Inheritance Probability
$b$	Budget Life Threshold
$\alpha$	Alpha Value Threshold
$ch$	Channels in Total for NCA
$gens$	Number of Generations for NCA to evolve
$act$	Activation on Board (applied to squash)
$ttr$	Expected Time to Run

Please refer the thesis report Table 4.4 for correct corresponding configuration of these Large Experiments. We outline total of 24 large experiments to download. You can also check tables below for corresponding configurations.

This table shows the experiments that are tried for long generations handpicked from small runs. Please refer thesis text for more info.

$exp.$	$W$	$H$	$init$	$ppp$	$ip$	$b$	$\alpha$	$ch$	$gens$	$act$	$ttr (h)$	Download Link
1	200	200	0.02	0.02	0.1	4	0.5	2	1000	sigmoid	42.8	Download exp1.tar
2	200	200	0.02	0.02	0.1	4	0	2	1000	tanh	43.9	Download exp2.tar
3	200	200	0.02	0.02	0.1	8	0.5	2	1000	sigmoid	44.2	Download exp3.tar
4	200	200	0.02	0.02	0.1	8	0	2	1000	tanh	40.0	Download exp4.tar
5	200	200	0.08	0.02	0.1	4	0.5	2	1000	sigmoid	40.9	Download exp5.tar
6	200	200	0.08	0.02	0.1	4	0	2	1000	tanh	38.6	Download exp6.tar
7	200	200	0.08	0.02	0.1	8	0.5	2	1000	sigmoid	43.1	Download exp7.tar
8	200	200	0.08	0.02	0.1	8	0	2	1000	tanh	42.3	Download exp8.tar
9	200	200	0.02	0.02	0.5	4	0.5	2	1000	sigmoid	39.2	Download exp9.tar
10	200	200	0.02	0.02	0.5	4	0	2	1000	tanh	43.3	Download exp10.tar
11	200	200	0.02	0.02	0.5	8	0.5	2	1000	sigmoid	40.7	Download exp11.tar
12	200	200	0.02	0.02	0.5	8	0	2	1000	tanh	44.5	Download exp12.tar
13	200	200	0.08	0.02	0.5	4	0.5	2	1000	sigmoid	38.8	Download exp13.tar
14	200	200	0.08	0.02	0.5	4	0	2	1000	tanh	42.4	Download exp14.tar
15	200	200	0.08	0.02	0.5	8	0.5	2	1000	sigmoid	40.1	Download exp15.tar
16	200	200	0.08	0.02	0.5	8	0	2	1000	tanh	41.7	Download exp16.tar
17	200	200	0.02	0.02	0.1	∞	0.5	2	1000	sigmoid	42.2	Download exp17.tar
18	200	200	0.02	0.02	0.1	∞	0	2	1000	tanh	39.6	Download exp18.tar
19	200	200	0.08	0.02	0.1	∞	0.5	2	1000	sigmoid	43.0	Download exp19.tar
20	200	200	0.08	0.02	0.1	∞	0	2	1000	tanh	44.4	Download exp20.tar
21	200	200	0.02	0.02	0.05	∞	0.5	2	1000	sigmoid	40.5	Download exp21.tar
22	200	200	0.02	0.02	0.05	∞	0	2	1000	tanh	38.9	Download exp22.tar
23	200	200	0.08	0.02	0.05	∞	0.5	2	1000	sigmoid	43.8	Download exp23.tar
24	200	200	0.08	0.02	0.05	∞	0	2	1000	tanh	41.1	Download exp24.tar

• All small run experiments can be downloaded from Download all_small_runs.tar

Cite this

To cite this repository:

@software{aiganca2023github,
  author = {Sanyam Jain and Stefano Nichele},
  title = {{AIGA}: Self-Replicating NCA with Coarse Grained Analysis},
  url = {http://github.com/s4nyam/MasterThesis},
  version = {1.0.1},
  year = {2023},
}

End Deliverables

This project involves dealing with large volumes of data. We divide the project and the simulation process into two sub-parts. The first part involves training the system parameters (also called genes or weights parameters of the neural network) and system environment parameters (or grid values). These parameters are optionally exported as pickle (package) files. In the second part, we load these weights parameters separately to perform the PD and GD analysis. This decision was made to balance the load on the computational resources and allow the use of these pickles at any time without running the complete system again.

The final deliverables include the system environment (NCA grid values and weight parameters if required) variables and the corresponding results of PD and GD tools. We release the following deliverables with this project:

1. The running code is available at the GitHub repository.

2. All tar files as part of compressed experimental data for small runs and large runs are available at Archive.org.

3. We also release high-quality resulting animations as a YouTube playlist.

4. A Google Colab Notebook is provided to try small runs on the go in small runtime environments at bit.ly/neuralCA.

Working pipeline Demo