Experiment 06 · constraint mechanisms @ 32×32 / 10 agents
Connectivity-mechanism shootout @ 32²
Pits the ways of enforcing the comm-graph constraint against each other at the target
scale: a soft λ₂ (algebraic-connectivity) penalty, a learned-Lagrangian constraint (RCPO), a
PID-Lagrangian controller, and a hard connectivity mask — each run both with explorer/relay roles
and with the homogeneous shared policy.
Key finding
With explorer/relay roles, the soft λ₂ penalty yields the highest coverage, followed
by the learned-Lagrangian (RCPO) and the hard connectivity mask, with the PID-Lagrangian trailing.
Roles beat the homogeneous policy under every matched mechanism — how the constraint is imposed and
whether the team has differentiated roles both matter, and they compound.