Transferable Mean-Field Equilibria for Dynamic Environments

Hu and Zhang (2024) introduce MF-OML, a breakthrough for computing Nash equilibria in large population games. However, it assumes a fixed game environment. Yet real-world systems (e.g., vehicular networks in Zhang et al., 2024) face dynamic changes (e.g., traffic shifts, new policies). This idea proposes transfer learning for mean-field games: pre-train an MF-OML agent in one environment and fine-tune it to new settings with minimal data. We’d leverage feature-based transfer (as in Zhang et al.’s VEC work) but apply it to occupation measures, allowing agents to adapt equilibrium policies across contexts (e.g., from highway to urban driving). This would address scalability gaps in Hu’s work while creating more resilient multi-agent systems. A key challenge: How do mean-field equilibria transfer across correlated environments? We’d explore this via theory and simulations in crowd dynamics (Talukdar et al., 2024).

References:

1. Neural Network-Based Game Theory for Scalable Offloading in Vehicular Edge Computing: A Transfer Learning Approach. Juan Zhang, Yulei Wu, Geyong Min, Keqin Li (2024). IEEE transactions on intelligent transportation systems (Print).
2. MF-OML: Online Mean-Field Reinforcement Learning with Occupation Measures for Large Population Games. Anran Hu, Junzi Zhang (2024). Applied Mathematics & Optimization.
3. Learning Crowd Motion Dynamics with Crowds. Bilas Talukdar, Yunhao Zhang, Tomer Weiss (2024). Proceedings of the ACM on Computer Graphics and Interactive Techniques.