HydraMARL
A foundational study of parameter sharing in cooperative multi-agent reinforcement learning. HydraMARL decomposes each agent's policy and value networks into a population-shared trunk and per-agent heads, applied symmetrically across actor and critic; a single continuous parameter spans the full spectrum from complete sharing to full independence. Systematic experiments on the VMAS benchmark suite identify the optimal sharing regime and characterise its sensitivity to task structure, degree of cooperation, and population size.
When many AI agents learn together, parts of their neural networks can be shared. We isolate how much sharing actually helps, and expose the trade-off between learning fast and learning to specialize.
