Deep Fictitious Play Trains Human-Like Driving Policies at Unsignalized Intersections
A recent study published on arXiv (2506.12283) presents an innovative framework for simulating vehicle interactions at unsignalized intersections through Deep Fictitious Play. This research redefines the interaction as a Potential Differential Game, utilizing cost function weights derived from naturalistic driving datasets to reflect a variety of driving behaviors. The framework ensures theoretical convergence to a Nash equilibrium, representing the inaugural application of Deep Fictitious Play for developing interactive driving policies. This study tackles the intricacies of game-theoretic dynamics at these intersections, enhancing previous methods that did not incorporate data-driven calibration.
Key facts
- The study is from arXiv:2506.12283.
- It uses Deep Fictitious Play to learn human-like driving policies.
- Vehicle interactions are modeled as a Differential Game, reformulated as a Potential Differential Game.
- Weights in the cost function are learned from naturalistic driving datasets.
- The framework guarantees convergence to a Nash equilibrium.
- This is the first study to train interactive driving policies using Deep Fictitious Play.
- The research focuses on unsignalized intersections.
- The approach captures diverse driving styles.
Entities
Institutions
- arXiv