Specify the interaction structure
This edition combines two MARL reports. State cooperation, competition, or mixed motives; team versus individual reward; local information; communication; parameter sharing; and test partners. PettingZoo and OpenSpiel are ecosystems/interfaces rather than single fixed evaluation tasks.
| Suite/platform | Structure | Required details |
|---|---|---|
| MPE/MPE2/PettingZoo | Lightweight particle and other tasks | AEC/Parallel API, scenario, wrappers |
| SMAC/SMACv2 | Partially observed team tactics | Maps, units, randomization, information |
| Multi-Agent MuJoCo | Cooperative continuous control | Joint partition, local inputs, agent count |
| Hanabi | Imperfect-information cooperation | Legal communication, hands, rules |
| Overcooked/RWARE | Coordination and planning | Layouts, horizon, partners, shaping |
| Melting Pot | Social interaction and unfamiliar partners | Substrates, background populations, tests |
| Google Research Football | Team control and competition | Scenario, controlled players, reward, opponents |
| OpenSpiel | Game-theoretic analysis | Game, information sets, best responses |
| MAgent2/MARLGrid | Large-scale or grid interaction | Population, observations, generation |
| Pommerman/RoboCup/Arena/Unity | Competitive/platform-specific tasks | Fork, rules, controls, external engines |
Fixed maps support regression tests, while generalization needs held-out scenarios or partners. Report cross-play separately from self-play.
Metric boundaries
Returns, success/win rates, and curves are a starting point. Communication needs message budgets and no-message/shuffled-message ablations. Social results require definitions of welfare and fairness.
For a suitable game,
This requires exact or approximate best responses and a stated approximation error. Exploitability normalization is implementation- and game-dependent. Cooperative win rate is not an equilibrium certificate. Elo/Glicko depend on the opponent population and can obscure nontransitivity.
Budgets and units
A joint environment transition may contain many agent actions. Agent steps, joint transitions, and rollout length differ. Agents sharing one team return are not independent replications. Record central-critic inputs, sharing, parallelism, updates, and complete wall time.
Interpretable comparisons
Fix maps, partner/opponent pools, and selection. A no-fresh-observation comparison can expose fixed-task open-loop shortcuts. Assess reward, unfamiliar-partner performance, and cost together.
Related reading: IPPO/COMA, statistics, and evaluation.