NAHT: N-Agent Ad Hoc Teamwork — experimental results

Training and evaluation logs, plus agent checkpoints, for the NAHT / POAM paper.

• Code: https://github.com/carolinewang01/naht — all training, evaluation, and plotting code. Every figure in the paper is reproducible from the logs here.

Also in this repo: naht_unctrl_agents_10-18-24.zip, a standalone 94 MB bundle of the uncontrolled-teammate agents, if you want only the pool and not the full logs.

Quick start

Download and extract; the archives restore the original directory layout, so the plotting notebooks in the code repo run against the extracted tree unmodified.

huggingface-cli download carolinewang01/naht --repo-type=dataset --local-dir naht_results
cd naht_results
for f in $(find . -name '*.tar.gz' -o -name '*.tar'); do tar xf "$f" -C .; done

Then point BASE_RES_PATH in the notebooks at the extracted directory.

Environments

Directory	Environment	Details
5v6/	SMAC	5m_vs_6m, 20M env steps
8v9/	SMAC	8m_vs_9m, 20M env steps
10v11/	SMAC	10m_vs_11m, 20M env steps
3sv5z/	SMAC	3s_vs_5z, 20M env steps
mpe-pp__ts=100_shape=0.01/	MPE	mpe:PredatorPrey-v0, 100-step episodes, reward shaping 0.01, 20M env steps
matrix-games__bit-3p-*/	Bit matrix game	bit-3p-jointactstate-v0 and bit-3p-nostate-v0, including the theory_ex encoder–decoder study

Experiment groups

One archive per <env>/<experiment-group>. Each is self-contained.

Archive	Contents
open_train	NAHT training runs — POAM (poam-*) and the IPPO-based ablations (ippo-*), including critic-masking (-cmask-) and team-composition (-teamcomp-) variants
ippo, mappo, iql, qmix, vdn (and *_ns no-parameter-sharing variants)	Self-play training runs for the uncontrolled-teammate pool
open_eval_best, open_eval_last	Cross-play evaluation under N-agent ad hoc teamwork
in_distr_eval	In-distribution teammate evaluation
ood_generalization, ood_gen_vp	Out-of-distribution teammate generalization
selfplay_mismatched_eval_best	Self-play and mismatched-partner evaluation
*__ed-tensors	Encoder–decoder tensor dumps from the poam_data-gathering runs; input to ed_vis_utils.load_data and the within-episode ED-loss figures. Stored uncompressed (.tar) since .npz is already compressed

Per-run layout

Every run directory follows the Sacred layout:

<experiment-group>/
  sacred/<run>/1/config.json     # full hyperparameter config
  sacred/<run>/1/run.json        # run metadata (start/stop, host, status)
  sacred/<run>/1/info.json       # logged scalars — what vis_utils.py reads
  tb_logs/<run>/events.out.*     # TensorBoard events — what the learning-curve code reads
  models/<run>/best/             # checkpoint with the best test return
  models/<run>/<final_step>/     # checkpoint at the end of training

Checkpoints are PyTorch .th state dicts (agent.th, critic.th, mixer.th where applicable, and for POAM additionally encoder.th / decoder.th), each with matching *_opt.th optimizer states.

Evaluation runs record the agents they load under trained_agents and unseen_agents in config.json; all of them load "load_step": "best". Those are absolute paths from the original cluster — rewrite the prefix to point at your local copy.

Citation

If you find our code or data useful, please cite:

@inproceedings{wang2024naht,
    title={N-Agent Ad Hoc Teamwork},
    author={Wang, Caroline and Rahman, Arrasy and Durugkar, Ishan and Liebman, Elad and Stone, Peter},
    booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2024}
}