jaxaht/tSNE-embeddings

JaxAHT behavioral embedding t-SNE

These are the classifier checkpoints, trajectory data, latents, figures and accuracy numbers behind the Behavioral Embedding t-SNE figures

Figure	Environment	Folder
Fig. 3	Overcooked-v1 Coordination Ring	coord_ring/
Fig. 9	LBF 7x7 (no levels)	lbf_7x7_nolevels/

Method

1. For every ordered pair (teammate T, best response to teammate S) in the held-out evaluation set, roll out episodes in which T plays with BR(S).
   • Coordination Ring: 18 teammates, so 18² = 324 pairs.
   • LBF: 23 teammates, so 23² = 529 pairs.
2. Train an LSTM trajectory classifier (LSTMTrajectoryEncoder → latent → linear head) to predict the pair identity. There are N² classes, using the teammate's (agent_0) observation sequence.
3. Encode the held-out test episodes. Run t-SNE (sklearn, perplexity 30, random_state=42) on the latents of the matched pairs only, where T plays its own best response BR(T).

Train/test split. Train and test episodes come from separate rollout calls with fresh RNG.

Teammate and best-response policies come from jaxaht/eval-teammates and jaxaht/eval-teammates-br.

Headline results

Accuracy is exact-pair (top-1 over N² classes). "Own BR" means the teammate is paired with its own best response; "Other BR" covers every other pairing.

Coordination Ring (Fig. 3): 324 classes, chance = 0.003

Split	All pairs	Own BR (18 pairs)	Other BR (306 pairs)
Test	0.233 (n=6,480)	0.564 (n=360)	0.213 (n=6,120)
Train	0.265 (n=15,552)	0.594 (n=864)	0.246 (n=14,688)

LBF 7x7 (Fig. 9): 529 classes, chance = 0.002

Split	All pairs	Own BR (23 pairs)	Other BR (506 pairs)
Test	0.423 (n=52,900)	0.438 (n=2,300)	0.423 (n=50,600)
Train	0.436 (n=994,193)	0.440 (n=77,004)	0.435 (n=917,189)

Per matched pair: accuracy when the teammate plays its own best response

Coordination Ring (Fig. 3)

Teammate	Train acc	Test acc	Gap	n_train	n_test
human_proxy	0.96	1.00	-0.04	48	20
brdiv_conf1_1	0.96	1.00	-0.04	48	20
independent_agent_0	0.96	0.90	+0.06	48	20
lbrdiv_conf_1_2	0.90	0.80	+0.10	48	20
lbrdiv_conf_1_1	0.79	0.75	+0.04	48	20
lbrdiv_conf_1_0	0.90	0.70	+0.20	48	20
brdiv_conf1_2	0.79	0.70	+0.09	48	20
plate_agent_0	0.40	0.65	-0.25	48	20
comedi_1_2	0.71	0.65	+0.06	48	20
comedi_1_0	0.50	0.60	-0.10	48	20
ippo_mlp_1	0.58	0.55	+0.03	48	20
comedi_1_3	0.33	0.40	-0.07	48	20
onion_agent_0	0.19	0.35	-0.16	48	20
ippo_mlp_3	0.38	0.35	+0.03	48	20
comedi_1_1	0.35	0.30	+0.05	48	20
brdiv_conf2_0	0.56	0.30	+0.26	48	20
comedi_1_4	0.38	0.15	+0.23	48	20
ippo_mlp_2	0.06	0.00	+0.06	48	20

LBF 7x7 (Fig. 9)

Teammate	Train acc	Test acc	Gap	n_train	n_test
comedi_1_1	0.91	0.87	+0.04	1031	100
entitled_agent	0.86	0.85	+0.01	2288	100
ippo_mlp_s2c0_2_0	0.66	0.66	+0.00	1659	100
brdiv_conf2_1	0.70	0.66	+0.04	1030	100
comedi_1_4	0.66	0.61	+0.05	5099	100
brdiv_conf1_2	0.64	0.60	+0.04	1433	100
lbrdiv_conf_1_2	0.59	0.59	-0.00	6080	100
lbrdiv_conf_1_0	0.53	0.59	-0.06	6120	100
comedi_1_3	0.59	0.59	-0.00	7163	100
brdiv_conf1_0	0.58	0.56	+0.02	1432	100
brdiv_conf2_0	0.55	0.51	+0.04	2289	100
brdiv_conf1_1	0.52	0.51	+0.01	1733	100
comedi_1_2	0.48	0.44	+0.04	5297	100
lbrdiv_conf_1_1	0.39	0.33	+0.06	6147	100
seq_agent_lexi	0.25	0.29	-0.04	1964	100
ippo_mlp_0	0.31	0.25	+0.06	6077	100
seq_agent_col	0.24	0.24	+0.00	2010	100
seq_agent_rcol	0.24	0.22	+0.02	2006	100
greedy_closest_teammate	0.23	0.21	+0.02	1945	100
comedi_1_0	0.14	0.20	-0.06	8235	100
seq_agent_farthest	0.20	0.14	+0.06	1956	100
seq_agent_rlexi	0.20	0.13	+0.07	1971	100
seq_agent_nearest	0.05	0.03	+0.02	2039	100

Run configuration

	Coordination Ring (Fig. 3)	LBF 7x7 (Fig. 9)
Classifier	original paper checkpoint (May 2026)	retrained Nov 2026: no saved checkpoint existed
Encoder hidden / latent	64 / 16	128 / 32
LR / epochs / batch	5e-4 / 200 / 1024	5e-4 / 200 / 2048
num_envs (train rollouts)	48	1024
Episode length	fixed 400 steps	variable, 5–100 steps
Train episodes per pair	exactly 48	1,024–8,235 (capped at 5,000 in training)
Test episodes per pair	20	100
Observation / episode	(400, 650) uint8	(≤100, obs_dim)

The number of train episodes per pair is num_envs × the number of episodes completed within one rollout, and incomplete tails are dropped.

• Coordination Ring episodes fill the whole rollout, so each env yields exactly 1 episode.
• LBF episodes are short, so each env yields many episodes in a 128-step rollout.

Layout

coord_ring/                      # Fig. 3 (Overcooked-v1 Coordination Ring)
  classifier/
    trajectory_classifier/       # orbax checkpoint
    accuracy_curve.png  loss_curve.png  training_losses.npy
  data/
    train_episodes.pkl           # 15,552 episodes (48 / pair)   3.8 GB
    val_episodes.pkl             #  6,480 episodes (20 / pair)   1.6 GB  <- test set
  latents/
    latents_test.pkl             # latents + predictions on val_episodes (used for the figure)
    latents_train.pkl            # same, on train_episodes
  figures/
    tsne_test.{png,pdf}          # the paper figure
    confusion_test.{png,pdf}
    tsne_train.{png,pdf}  confusion_train.{png,pdf}
  results/
    summary.json                 # overall / own-BR / other-BR accuracy, train & test
    per_pair_test.csv  per_pair_train.csv
lbf_7x7_nolevels/                # Fig. 9 (LBF 7x7, no levels); same structure
  data/train_episodes.pkl        # 994,193 episodes   3.0 GB
  data/val_episodes.pkl          #  52,900 episodes (100 / pair)  201 MB  <- test set

latents_*.pkl is a dict with these keys:

• latents_dict: {matched_pair_label: [latent, ...]}, the matched pairs only.
• predictions, all_true_labels: per episode, over all pairs.
• label_to_idx, idx_to_label, num_classes.

Pair labels have the form <BR source agent>_br_for_<teammate>. Separators are inconsistent across names (brdiv-conf1_1 vs brdiv_conf1_1), so use the latents_dict keys or the matched column in the CSVs to find own-BR pairs.