Events as Triggers for Behavioral Diversity in Multi-Agent Reinforcement Learning

Abstract

A multi-agent reinforcement learning framework in which events, such as a teammate dropping out or a door opening, trigger behavioral transitions. An event-driven hypernetwork generates LoRA modules over a shared team policy, decoupling behavior from agent identity, while diversity is held at a target value of a new metric, Neural Manifold Diversity, provably without hindering reward maximization. In VMAS, the framework generalizes zero-shot to unseen team sizes and capability values and, unlike the baselines, solves a task requiring sequential role reassignment.

Paper