Quickstart

Setup virtual environment

XLRON uses uv to manage dependencies. Clone the repository and install:

git clone https://github.com/micdoh/XLRON.git
cd XLRON

# Install with uv (recommended)
uv sync

# Install dev dependencies
uv sync --group dev

# If running on GPU
uv sync --group gpu

# If running on TPU
uv sync --group tpu

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 xlron.train.train entry point. For example, to recreate the training of DeepRMSA:

uv run python -m xlron.train.train --env_type=deeprmsa --continuous_operation --load=250 --k=5 --topology_name=nsfnet_deeprmsa_directed --link_resources=100 --mean_service_holding_time=25 --truncate_holding_time --ROLLOUT_LENGTH=100 --NUM_LAYERS 5 --NUM_UNITS 128 --NUM_ENVS 16 --TOTAL_TIMESTEPS 5000000 --ENV_WARMUP_STEPS 3000 --LR 5e-5 --LR_SCHEDULE linear --UPDATE_EPOCHS 10 --GAE_LAMBDA 0.9 --GAMMA 0.95 --ACTION_MASKING

To evaluate a heuristic (e.g. KSP-FF) instead of training an agent:

uv run python -m xlron.train.train --env_type=rmsa --topology_name=nsfnet_deeprmsa_directed --link_resources=100 --k=50 --load=250 --continuous_operation --ENV_WARMUP_STEPS=3000 --TOTAL_TIMESTEPS=20000000 --NUM_ENVS=2000 --EVAL_HEURISTIC --path_heuristic=ksp_ff

You can save models with --SAVE_MODEL to the directory specified by --MODEL_PATH, and evaluate a saved model with --EVAL_MODEL --MODEL_PATH <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 --EPISODE_DATA_OUTPUT_FILE.