Contact: simon.hirlaender(at)sbg.ac.at

Official repository for the FERMI Free-Electron Laser (FEL) paper, utilizing model-based and model-free reinforcement learning methods to solve complex particle accelerator operation problems. This work demonstrates the practical application of deep RL for intensity optimization, comparing the sample-efficient AE-DYNA (model-based) with the high-performing NAF2 (model-free) algorithms.

conda create -n fermi_rl python=3.8 conda activate fermi_rl pip install -r requirements.txt

# NAF2 (Model-Free) python src/run_naf2.py # AE-DYNA (Model-Based) python src/AE_Dyna_Tensorflow_2.py

NAF2 Training	NAF2 Convergence

AE-DYNA Training	AE-DYNA Verification

Noise Robustness	Sample Efficiency (NAF vs AE-DYNA)

. ├── src/ # Python source code │ ├── run_naf2.py # NAF2 agent (TF2) │ ├── AE_Dyna_Tensorflow_2.py # AE-DYNA agent (TF2) │ └── AEDYNA.py # AE-DYNA agent (TF1.15 legacy) ├── paper/ # LaTeX source and figures │ ├── main.tex │ └── Figures/ ├── data/ # Experimental data └── requirements.txt 

If you use this work, please cite:

@software{hirlaender_fermi_rl, author = {Hirlaender, Simon and Bruchon, Niky}, title = {FERMI RL Paper Code}, year = 2020, publisher = {Zenodo}, doi = {10.5281/zenodo.4348989}, url = {https://doi.org/10.5281/zenodo.4348989} }