New neural framework reverse-engineers crowd dynamics from raw data
Hidden preferences and constraints in population games can now be learned automatically.
Deep Dive
A neural network framework learns hidden parameters of finite-state Mean Field Games from observed population dynamics. Using implicit differentiation to backpropagate through game equilibrium, the method works without needing agent actions or rewards. It was tested on synthetic benchmarks and real urban mobility data.
Key Points
- Framework uses implicit differentiation to backpropagate through MFG equilibrium without observing agent actions or rewards
- Supports state- and time-dependent parameter schedules, enabling trajectory-wise calibration
- Validated on four systems, including real urban mobility datasets, demonstrating practical applicability
Why It Matters
Enables data-driven modeling of complex strategic systems like traffic, markets, and epidemics without hand‑crafting hidden parameters.