Parameter Sharing / specialisation
Reassessing the Role of Parameter Sharing for Specialisation in Multi-Agent Reinforcement Learning. In a team, agents usually share one network, but sooner or later they need to act differently, so many methods give each agent a little capacity of its own and report gains. We ask whether those gains come from the mechanism or simply from that agent-specific capacity, which under a fixed parameter budget has a price: every parameter owned by one agent is one the others cannot use. Against a plain reference, a shared trunk with per-agent heads, we compare about twenty methods at matched budgets, in theory and experiment, to see how much private capacity a task really needs and where it belongs.
Many multi-agent methods let each agent keep part of the network to itself, and report gains. We ask whether the gains come from the mechanism or just from that private capacity, which on a fixed budget always costs the shared part.
