Dynamic Relational Architectures: Real-Time Evolution of Social Networks in Learning Multi-Agent Systems

Haeri (2022) introduced Reward-Sharing Relational Networks (RSRN), but these networks are statically defined a priori (e.g., self-interested, communitarian). What if, instead, the relational network—the graph of who cares about whom—could adaptively evolve based on ongoing experience, local rewards, or meta-learning? This research would develop algorithms where agents not only learn their own policies, but also adjust the weights and connections in their social network in response to observed collective outcomes, perhaps even pruning or forging new "social" ties dynamically. This meta-level of emergence (emergent social structure, not just emergent behavior) could reveal new principles of organization, and might model phenomena like the spontaneous formation of cliques, alliances, or hierarchies—going beyond both fixed-network MARL and existing relational frameworks.

References:

1. Reward-Sharing Relational Networks in Multi-Agent Reinforcement Learning as a Framework for Emergent Behavior. H. Haeri (2022). Adaptive Agents and Multi-Agent Systems.
2. Reward-Sharing Relational Networks in Multi-Agent Reinforcement Learning as a Framework for Emergent Behavior. H. Haeri (2022). Adaptive Agents and Multi-Agent Systems.
3. Reward-Sharing Relational Networks in Multi-Agent Reinforcement Learning as a Framework for Emergent Behavior. H. Haeri (2022). Adaptive Agents and Multi-Agent Systems.