We compare Lamarckian evolution framework with Darwinian evolution framework in a mujoco based wrapper called Revolve2. The release version of Revolve2 used in this project is v0.3.8-beta1 (https://github.com/ci-group/revolve2/releases/tag/v0.3.8-beta1).
Both frameworks were tested with two tasks (panoramic rotation and point navigation). Parameters used in the experiments are:
pop_size=50,
offspring_size=25,
nr_generations=30,
learner==RevDE,
learning_trials=280,
simulation_time=30 secs for rotation task,
simulation_time=40 secs for point navigation task.
In addition, two control experiments have been conducted. The first one (located under the "fixed_body" branch) involves utilizing the initial 50 robots' bodies from the main experiments, while the second one (located under the "random_body" branch) entails employing random bodies at each generation.
Steps to install:
1. git clone git@github.com:onerachel/Lamarckian_vs_Darwinian.git
2. cd Lamarckian_vs_Darwinian
3. virtualenv .venv
4. source .venv/bin/activate
5. ./dev_requirements.sh
To run experiments, e.g. darwinian_point_navigation:
python darwinian_evolution/optimize.py
To show the simulation, add --visualize:
python darwinian_evolution/optimize.py --visualize
To restart from the last optimization checkpoint, add --from_checkpoint:
python darwinian_evolution/optimize.py --from_checkpoint
To plot fitness:
python darwinian_evolution/plot_fitness.py
To check the best robot wrt the fitness:
cd darwinian_evoluation
python rerun_best.py
To check the best robot wrt the fitness and save the video:
cd darwinian_evoluation
python rerun_best.py -r <OUTPUT-DIR>