Selection for short-term empowerment accelerates the evolution of homeostatic neural cellular automata

This repository contains source code for the GECCO'23 (Genetic and Evolutionary Computation Conference 2023) paper

Caitlin Grasso and Josh Bongard (2023). Selection for short-term empowerment accelerates the evolution of homeostatic neural cellular automata.

Note: The arxiv version of this paper contains corrections to the methods as well as updated figures with new data from evolution with an increased population size. Please refer to arxiv (linked above) for the most up-to-date version of the paper.

Structure

• main.py - launches an evolutionary algorithm that performs bi- or tri-objective search of NCA capable of morphogenesis. Specify the parameters of the EA as command line input. For example:

python main.py --run=1 --gens=2000 --popsize=400 --target=square --objective1=error --objective2=MI --k=1

      See notes below regarding the objective names.

• constants.py - allows you to set certain constants such as the number of cellular automata (CA) iterations, the CA grid size, # EA generations, EA population size, and more. Parameters set as input to main.py will overwrite the values in constants.py.

• config.py - holds the target shapes. To create and test additional target shapes, add them to this file.

• visualize_sim.ipynb - python notebook to visualize a CA simulation and save out a video.

• analysis/sort_by_objective.py - sorts results from the EA and returns the best (or worst) performing NCA on the particular objective given as input.

• analysis/ - directory containing all code to analyze various EA runs and produce the results/plots in the paper.

  • In particular, analysis/plot_avg_fitness_curves.py is using to generate the loss plots seen throughout the paper.

• additional_objectives.py contain the additional objectives that were tested in Figure 9 of the paper.

Experiments

1. Morphogenesis (Fig. 2)

History length (k) sweep.

Run main.py for each of the evolutionary variations as follows:

Bi-loss

python main.py --runs=[0...35] --gens=2000 --popsize=400 --target=square --objective1=error

Tri-loss

python main.py --runs=[0...35] --gens=2000 --popsize=400 --target=square --objective1=error_phase1 --objective2=error_phase2

Tri-loss-empowerment (with varying k)

python main.py --runs=[0...35] --gens=2000 --popsize=400 --target=square --objective1=error --objective2=MI --k=[1,5,10,17,25,32,40,45]

Evolution is run for each value in the range or set of values indicated by the brackets. Thus, for the bi-loss and tri-loss trials, 35 calls to main.py are made. For the tri-loss-empowerment trials, 35x8=280 calls to main.py are made.

---

2. Homeostasis (Figs 3 & 4)

• analysis/plot_extended_sim_stability_all.py produces results related to stability if Figures 3 & 4.

• analysis/morphogenetic_homeostatic_characteristics.py generates the bar plots in Figure 4 relating to # connected components, cell state transiency, and proportion of cells on boundary

---

3. Generalization (Figs 5 & 6)

1. Generalization to different shapes with empowerment pre-trained NCA (Fig 5).

• finetune.py is similar to main.py but "continues" evolution from a checkpointed generation (i.e. evolution does not start with a random population)
• To produce the generalization results in Figure 5, the runs described in part 1 of this section must already be performed (for the tri-loss-empowerment variation only k=1 needs to be run). finetune.py can then be edited to point to a checkpointed file (by setting the PATH_TO_CHECKPOINT variable) from one of these runs and evolution can be continued.
• The --target option can be set to change the target shape to the triangle, x, or biped.

2. Generalization to larger grid (Fig 6).

• Experiments described in part 1 of this section were repeated but with the following changes to constants.py

GRID_SIZE = 50
ITERATIONS = 100

• Only three values of history length were tested: k=[1,45,90]

Notes:

1. python3 required to run the code in this repository.
2. error is the name of the loss metric and MI is the name of the empowerment metric. Thus, runs labeled error_MI in the code are called tri-loss-empowerment runs in the paper (tri-objective search with the first objective as age, the second as loss, and the third as empowerment). A table mapping the objective syntax to the EA variations described in the paper is below.

EA Variation	Objective 1	Objective 2
bi-loss	error	-
tri-loss	error_phase1	error_phase2
tri-loss-empowerment	error	MI

4. The run parameter input to main.py is the random seed for the evolutionary run.