Learning to communicate

A JAX-based slotted-channel simulator for multi-agent medium access control, together with the deep-RL agent it was built for and a set of reference baselines from the literature.

The proposed method is a DDQN agent over an MLP Q-network, trained fully online and without coordination between stations. Everything else in ltc/baselines/ is a comparison point.

Install

git clone https://github.com/mldr-devs/ltc.git
cd ltc                     # requires Python >= 3.12
pip install -e .

Optional extras: .[gpu] for CUDA, .[symbolic] for the distillation pipeline, .[analysis] for the sensitivity analysis.

Run

# the method: 10 DDQN stations, 50 x 2000 slots
python -m ltc.run --n 10 --n_epochs 50 --n_steps 2000 --seed 42

Each run writes a compressed history file (history_<n>_<n_final>_<seed>_<commit>.pkl.lz4) holding the final agent states, the per-slot outputs, and the invocation arguments. The commit hash is recorded automatically, and the run refuses to start with uncommitted tracked changes unless you pass --skip_git_check.

Cite

@article{ltc,
  title={{Learning to Communicate}},
  author={Szczech, Kamil and Wojnar, Maksymilian and Rusek, Krzysztof and Kosek-Szott, Katarzyna and Szott, Szymon},
  year={2026}
}