Framework combines global exploration intensity control with per-agent signal quality allocation for cooperative multi-agent reinforcement learning

Researchers address two challenges in cooperative multi-agent reinforcement learning (MARL): adapting exploration intensity globally during training, and allocating exploration budget across agents based on the reliability of their intrinsic reward signals.

The framework uses a return-conditioned sigmoid schedule (RCB) for global intensity control and a per-agent Reward Signal Quality (RSQ) metric. Successor Distance (SD), a quasimetric intrinsic reward, produces distinguishable per-agent signal quality and includes convergence and ordering preservation guarantees.

Method achieves top-tier returns across all seven cooperative benchmarks tested (MPE, SMAX, MABrax).