derived/reports/report.md

MARL Experiment Summary

Generated: 2026-07-02T00:51:16.148597+00:00 Tables: runs/tables/mouse-run-run-1

Valid Pairs

Task	Valid pairs
non_social	10
social	10

Training Metrics (final update, mean ± std across pairs)

Metric	non_social (n=10)	social (n=10)
collisions per episode	2.02 ± 0.745	4.758 ± 6.29
chaser return	6.425 ± 0.352	-1.839 ± 7.419
explorer return	61.896 ± 9.459	23.844 ± 21.326
chaser partner vision	0 ± 0	0.326 ± 0.228
explorer partner vision	0 ± 0	0.326 ± 0.228
chaser new fields	82.535 ± 1.845	15.063 ± 4.095
explorer new fields	74.988 ± 6.285	51.49 ± 12.231
final distance	5.044 ± 0.426	5.429 ± 1.504
policy loss	-0.000564 ± 0.000452	0.000734 ± 0.005
value loss	5.034 ± 0.193	7.117 ± 0.718
entropy	0.25 ± 0.071	0.631 ± 0.166
approx kl	0.000188 ± 0.000384	0.004 ± 0.006
chaser grad norm	0.296 ± 0.005	2.553 ± 1.475
explorer grad norm	0.313 ± 0.059	1.061 ± 0.595
max param abs	1.782 ± 0.497	2.376 ± 0.987

Random-Opponent Evaluation

Opponent: random_chaser

Metric	non_social (n=10)	social (n=10)
collisions per episode	2.872 ± 0.316	0.572 ± 0.192
chaser return	-3.292 ± 0.037	-2.605 ± 0.208
explorer return	61.532 ± 9.081	17.34 ± 18.824
chaser partner vision	0 ± 0	0.266 ± 0.031
explorer partner vision	0 ± 0	0.266 ± 0.031
chaser new fields	33.841 ± 0.19	33.939 ± 0.097
explorer new fields	74.874 ± 6.064	45.239 ± 12.602
final distance	4.862 ± 0.21	5.437 ± 0.242
average distance	4.996 ± 0.075	5.584 ± 0.193
degenerate fraction	0.882 ± 0.019	0.925 ± 0.039

Opponent: random_explorer

Metric	non_social (n=10)	social (n=10)
collisions per episode	3.451 ± 0.325	8.976 ± 5.699
chaser return	6.316 ± 0.366	3.174 ± 7.155
explorer return	-1.299 ± 0.334	-10.191 ± 6.452
chaser partner vision	0 ± 0	0.605 ± 0.185
explorer partner vision	0 ± 0	0.605 ± 0.185
chaser new fields	82.564 ± 1.901	17.55 ± 5.312
explorer new fields	32.97 ± 0.246	30.951 ± 1.595
final distance	5.005 ± 0.188	4.094 ± 1.119
average distance	5.191 ± 0.038	3.777 ± 0.992
degenerate fraction	0.845 ± 0.011	0.928 ± 0.059

Failed evaluations: 0

Figures

training curves

Training curves, mean ± std across 20 valid pairs.

evaluation comparison

Evaluation metrics over 40 successful evaluations, grouped by opponent mode and task.

Degenerate Episode Exclusions

Excluded episodes	Episodes with records	Records
82	225	9

Failed Attempts

• count: 0