This paper introduces a meta‑multi‑agent reinforcement learning (meta‑MARL) framework that enables fast adaptation of interactive policies in multi‑agent systems. Meta‑reinforcement learning employs a bi‑level optimization to let agents quickly adjust to new tasks or environments, yet existing approaches focus on single‑agent settings and cannot be directly extended to multi‑agent scenarios where tasks are shaped by both the environment and strategic interactions among agents. To tackle this, we model multi‑agent reinforcement learning problems as Markov games and train policies that can rapidly transfer across a distribution of such games. A new solution concept, meta‑Nash equilibrium (meta‑NE), is defined as the target of the meta‑MARL problem. We prove that, under mild smoothness assumptions, a meta‑NE coincides with a stationary point of a gradient‑play‑based meta‑learning algorithm, providing convergence guarantees. Experiments on interactive autonomous‑driving tasks compare against pretrained MARL baselines and show that meta‑MARL achieves comparable or superior performance with far fewer interaction steps, confirming the effectiveness of the proposed framework.
These results illustrate a natural synergy between meta‑learning and multi‑agent game theory, offering a practical route for systems that require rapid policy updates such as vehicle‑to‑infrastructure coordination and collaborative robotics.
Review