Baselines¶

We trained seven MARL algorithms on every scenario with the bundled Sample Factory trainer: IPPO, MAPPO, HAPPO, IDQN, VDN, QMIX and QPLEX. Scores are each scenario's true_objective. See the paper for the exact protocol.

Baseline results heatmap

Learning curves¶

Baseline learning curves

Scaling the team¶

Armory Siege with 2 to 8 agents (QMIX):

Agent scaling: learning

Agent scaling: throughput

To rerun these experiments, see Reproducing the paper.