This repository is linked to the following publication :
Q. Vacher, S. Kelly, A. Naqvi, N. Beuve, T. Djavaherpour, M. Dardaillon, K. Desnos. MAPLE: Multi-Action Programs through Linear Evolution for Continuous Multi-Action Reinforcement Learning
This repository contain:
- The code and script to reproduce the results.
- The experimental results to generate the figures and tables in the paper.
To execute the code, you will need first to download the repository. Then launch the script to install the libraries of MuJoCo and Gegelati and compile the project.
git clone --recurse-submodules https://github.com/gegelati/GECCO2025.git
cd GECCO2025
sudo scriptLibraries.sh
Now you can build the projet.
cd gegelati-apps/mujoco/ && mkdir bin && cd bin
cmake -DCMAKE_BUILD_TYPE=Release .. && cmake --build . -j
cd ../..
Two nootebooks are available:
- runTraining.ipynb To launch trainings with either MAPLE, MATPG or TPG on any MuJoCo environment
- results.ipynb To look or generate the different results in the paper For runTraining.ipynb, or for the last section of results.ipynb, be sure to have the same OS for the compiler of the project and for the notebook runner.
├─ GECCO2025 # Root folder
│ │
│ │
│ ├─ exported_results
│ │
│ ├─ gegelati-apps/mujoco
│ │
│ ├─ logs # Experimental data
│ │ │
│ │ │
│ │ ├─ MAPLE
│ │ │ │
│ │ │ ├─ ant
│ │ │ │ │
│ │ │ │ ├─ codeGen
│ │ │ │ │
│ │ │ │ ├─ bestPolicyStats.0.p0.ant.md
│ │ │ │ │
│ │ │ │ ├─ exported_params.ant.p0.json
│ │ │ │ │
│ │ │ │ ├─ out_best_stats.0.p0.ant.md
│ │ │ │ │
│ │ │ │ ├─ out_best.0.p0.ant.dot
│ │ │ │ │
│ │ │ │ ├─ stateAndActionData_out_best.0.p0.42.csv
│ │ │ │ │
│ │ │ ├─ half_cheetah
│ │ │ │
│ │ │ ├─ hopper
│ │ │ │
│ │ │ ├─ humanoid
│ │ │ │
│ │ │ ├─ inverted_double_pendulum
│ │ │ │
│ │ │ ├─ walker2d
│ │ │ │
│ │ ├─ MATPG
│ │ │
│ │ ├─ TPG
│ │ │
│ │ │
│ ├─ videoResults
│ │
│ ├─ README.md
| |
│ ├─ results.ipynb
| |
│ ├─ mann_whitney_maple.csv
│ │
│ ├─ scriptLibraries.sh
│ │
│ ├─ sumup_rl_results.csv
│ │