Quickstart

Setup virtual environment

We recommend using poetry to manage dependencies. To install poetry, run:

pip install poetry

Then, clone the repository and install the dependencies:

git clone
cd XLRON
poetry install

If you prefer, you can use requirements.txt (not recommended as it may not be up-to-date):

pip install -r requirements.txt

N.B. if you want to run on GPU or TPU, then you need to edit the pyproject.toml file to include the appropriate accelerator-specific dependencies, i.e. comment/uncomment the relevant lines below, then follow the steps above:

jax = {extras = ["tpu"], version = "^0.4.11"}  # for Google Cloud TPU
jax = {extras = ["cuda11_pip"], version = "^0.4.11"} # or cuda12_pip depending on drivers
jax = {extras = ["cpu"], version = "^0.4.11"}  # for CPU-only

Running experiments

Every experiment in XLRON is defined through command-line options. No need to edit the source code. See configuration options for a full list of options.

To run an experiment, use the train.py script. For example, to recreate the training of DeepRMSA:

poetry run python train.py --env_type=deeprmsa --env_type=deeprmsa --continuous_operation --load=250 --k=5 --topology_name=nsfnet_deeprmsa --link_resources=100 --max_requests=1e3 --max_timesteps=1e3 --mean_service_holding_time=25 --ROLLOUT_LENGTH=100 --continuous_operation --NUM_LAYERS 5 --NUM_UNITS 128 --NUM_ENVS 16 --TOTAL_TIMESTEPS 5000000 --ENV_WARMUP_STEPS 3000 --LR 5e-5 --WARMUP_PEAK_MULTIPLIER 2 --LR_SCHEDULE linear --UPDATE_EPOCHS 10 --GAE_LAMBDA 0.9 --GAMMA 0.95 --ACTION_MASKING

Launch power optimisation using closed-form GN model RMSA environment (experimental feature):

poetry run python train.py --env_type=rsa_gn_model --PROJECT LAUNCH_POWER --ROLLOUT_LENGTH=50 --NUM_LAYERS 3 --NUM_UNITS 128 --load=100 --k=5 --weight=weight --topology_name=nsfnet_deeprmsa_directed --link_resources=115 --max_requests=10 --max_timesteps=10 --mean_service_holding_time=25 --continuous_operation --ENV_WARMUP_STEPS=0 --TOTAL_TIMESTEPS 20000 --NUM_ENVS 200 --launch_power_type=rl --interband_gap=100 --values_bw=400,600,800,1200 --guardband=0 --coherent --reward_type=bitrate --snr_margin=0.01 --slot_size=100 --max_power=3 --min_power=-1 --VISIBLE_DEVICES=1 --PLOTTING --DOWNSAMPLE_FACTOR=1 --LR 1e-4 --LR_SCHEDULE warmup_cosine --WARMUP_PEAK_MULTIPLIER 2 --GAE_LAMBDA 0.9 --GAMMA 0.9  --EVAL_MODEL --LOAD_MODEL --MODEL_PATH /home/XLRON/models/BEST_LP --WANDB --DATA_OUTPUT_FILE /home/XLRON/data/launch_power_train_out.csv

You can save models with --SAVE_MODEL to the directory specified by --MODEL_PATH. You can log data to Weights and Biases (see Understanding XLRON for more details) with --WANDB. You can also log episode end data to a CSV file with --DATA_OUTPUT_FILE.