AAMAS Experiments
Use this before submission when the empirical or simulation story is not yet locked. At AAMAS
the experiment exists to test the interaction claim, not to top a benchmark.
Experiment audit
- Map each empirical claim to a game, a self-play run, a population sweep, an ablation, or a
deviation test.
- Choose opponents deliberately: self-play alone rarely suffices; include held-out opponents,
population sets, or classical strategies as the claim requires.
- Separate simulations that validate a solution concept (where the equilibrium is known) from
real or applied studies that show practical multiagent behavior.
- Report uncertainty for stochastic results over both seeds and opponents: standard errors,
confidence intervals, or paired tests.
- Report the environment, number of agents, training regime, evaluation protocol, metrics,
hyperparameter ranges, chosen settings, seeds, hardware, software versions, and runtime.
- Add ablations for the interaction mechanism (communication, reward sharing, the payment rule),
not just cosmetic variants.
- Audit for the mismatch between the strategic claim and the setup: an equilibrium claim tested
against only one fixed opponent, or a cooperation claim that hides a reward-shaping constant.
What experiments are for at this venue
- The strongest design shows the interaction under stress: agents that can deviate, opponents
the method did not train against, and populations that vary in size or composition.
- One experiment that lets agents try to exploit the mechanism and fails to profit is worth more
than five extra environments where nothing strategic is tested.
- Reviewers, often game theorists, check whether the metric matches the claim: convergence to a
named solution concept, exploitability, social welfare, or regret - not just episodic return.
Interaction-validation design table
| Interaction claim |
Matching experiment |
Reject pattern avoided |
| Converges to equilibrium |
Convergence/exploitability curve under simultaneous adaptation |
"Equilibrium asserted, never measured" |
| Mechanism is truthful |
Strategic-deviation test: an agent tries to misreport |
"Truthfulness proved, never stress-tested" |
| Beats other agents |
Round-robin vs held-out opponents and a population |
"Self-play only" |
| Emergent cooperation |
Sweep over reward/opponent settings with variance |
"One seed, one setting, one story" |
Vignette: a coordination-protocol study
Suppose the paper claims a learned protocol raises cooperation in a repeated public-goods game.
The matching plan: sweep group size and defector fraction for cooperation curves, add held-out
opponents that never appeared in training, and inject a free-rider agent to measure whether it
profits - every panel tied to a numbered claim or definition.
Statistical reporting floor
- Seeds and replication counts for every stochastic curve; captions must state whether bands are
standard errors, confidence intervals, or quantiles, and how many opponents were averaged.
- Report the compute actually consumed by self-play, not vague feasibility language.
Output format
[Experiment readiness] strong / adequate / weak
[Claim -> evidence map] <claim: game / self-play / population / deviation test>
[Missing interaction evidence] <opponents / deviation test / seeds / metric>
[Reproducibility gaps] <hyperparameters / compute / env / seeds>
[Decision-critical next run] <one experiment or simulation>
1---2name: aamas-experiments3description: Use when designing or auditing AAMAS experiments - self-play and population-based training, opponent selection, equilibrium and regret metrics, game-theoretic simulations, ablations, seeds, hyperparameters, compute, and claim-to-evidence fit - with emphasis on experiments that probe the interaction rather than chase a single-agent leaderboard.4---56# AAMAS Experiments78Use this before submission when the empirical or simulation story is not yet locked. At AAMAS9the experiment exists to test the *interaction* claim, not to top a benchmark.1011## Experiment audit1213- Map each empirical claim to a game, a self-play run, a population sweep, an ablation, or a14 deviation test.15- Choose opponents deliberately: self-play alone rarely suffices; include held-out opponents,16 population sets, or classical strategies as the claim requires.17- Separate simulations that validate a solution concept (where the equilibrium is known) from18 real or applied studies that show practical multiagent behavior.19- Report uncertainty for stochastic results over both seeds and opponents: standard errors,20 confidence intervals, or paired tests.21- Report the environment, number of agents, training regime, evaluation protocol, metrics,22 hyperparameter ranges, chosen settings, seeds, hardware, software versions, and runtime.23- Add ablations for the interaction mechanism (communication, reward sharing, the payment rule),24 not just cosmetic variants.25- Audit for the mismatch between the strategic claim and the setup: an equilibrium claim tested26 against only one fixed opponent, or a cooperation claim that hides a reward-shaping constant.2728## What experiments are for at this venue2930- The strongest design shows the interaction under stress: agents that *can* deviate, opponents31 the method did not train against, and populations that vary in size or composition.32- One experiment that lets agents try to exploit the mechanism and fails to profit is worth more33 than five extra environments where nothing strategic is tested.34- Reviewers, often game theorists, check whether the metric matches the claim: convergence to a35 named solution concept, exploitability, social welfare, or regret - not just episodic return.3637## Interaction-validation design table3839| Interaction claim | Matching experiment | Reject pattern avoided |40|---|---|---|41| Converges to equilibrium | Convergence/exploitability curve under simultaneous adaptation | "Equilibrium asserted, never measured" |42| Mechanism is truthful | Strategic-deviation test: an agent tries to misreport | "Truthfulness proved, never stress-tested" |43| Beats other agents | Round-robin vs held-out opponents and a population | "Self-play only" |44| Emergent cooperation | Sweep over reward/opponent settings with variance | "One seed, one setting, one story" |4546## Vignette: a coordination-protocol study4748Suppose the paper claims a learned protocol raises cooperation in a repeated public-goods game.49The matching plan: sweep group size and defector fraction for cooperation curves, add held-out50opponents that never appeared in training, and inject a free-rider agent to measure whether it51profits - every panel tied to a numbered claim or definition.5253## Statistical reporting floor5455- Seeds and replication counts for every stochastic curve; captions must state whether bands are56 standard errors, confidence intervals, or quantiles, and how many opponents were averaged.57- Report the compute actually consumed by self-play, not vague feasibility language.5859## Output format6061```text62[Experiment readiness] strong / adequate / weak63[Claim -> evidence map] <claim: game / self-play / population / deviation test>64[Missing interaction evidence] <opponents / deviation test / seeds / metric>65[Reproducibility gaps] <hyperparameters / compute / env / seeds>66[Decision-critical next run] <one experiment or simulation>67```
brycewang-stanford