onerachel / lamarckism_environments

Lamarckian Inheritance Improves Robot Evolution in Dynamic Environments

We tested the Lamarckian system both in the real world and in a MuJoCo-based wrapper called Revolve2. The release version of Revolve2 used in this project is v0.3.8-beta1, available at https://github.com/ci-group/revolve2/releases/tag/v0.3.8-beta1. The setup for the physical experiments can be found here: https://github.com/onerachel/physical_robot_experiments.

Parameters used in the experiments are:

ENVIRONMENT = "FLAT" or "RUGGED" # type of environment
CHANGE_FREQUENCY = 0 or 2 or 5   # frequency of environmental changes during an evolutionary run
pop_size=50,
offspring_size=25,
nr_generations=30,
learner==RevDE,
learning_trials=280

Installation

Steps to install:

1. git clone https://github.com/onerachel/Lamarckism_Environments.git
2. cd Lamarckism_Environments
3. virtualenv -p=python3.8 .venv
4. source .venv/bin/activate
5. ./dev_requirements.sh

In case you have installation errors wrt multineat, find where the cereal library is installed, e.g.: find / -name "cereal*", then tell CPPNWIN where to look for cereal manually, e.g.: export CPATH=/usr/local/include

Run experiments

To run experiments

cd lamarckian_evolution
python optimize.py

To show the simulation, add --visualize:

python optimize.py --visualize

To restart from the last optimization checkpoint, simply rerun:

python optimize.py

To plot fitness:

python plot_fitness.py

To check the best robot wrt the fitness:

python rerun_best.py

To check the best robot wrt the fitness and save the video:

python rerun_best.py -r <OUTPUT-DIR>

Highlighted results

Documentation

ci-group.github.io/revolve2