Leadership Inference for Multi-Agent Interactions

Hamzah Khan, David Fridovich-Keil

Introduction

In complex multi-agent environments, such as highway driving, agents must constantly negotiate for space and priority. Traditional heuristics often rely on spatial cues—for instance, assuming the leading vehicle is the "leader." However, such intuitions are frequently insufficient. An aggressive driver tailgating a cautious one might be the true "leader" forcing a pace, or simply a "follower" reacting to the front car's behavior.

Accurately deciphering these latent dynamics is critical for autonomous vehicles (AVs) to predict intent and plan safe, socially-compliant maneuvers. This paper introduces a formal framework to identify leadership in real-time by modeling it as a structural property of a game-theoretic interaction, rather than a mere spatial relationship.

Methodology

The proposed framework relies on two primary technical innovations: Stackelberg Iterative Linear-Quadratic Games (SILQGames) and the Stackelberg Leadership Filter (SLF).

SILQGames

While standard game theory often seeks Nash equilibria, leadership is more naturally modeled through Stackelberg equilibria. In this model, a "leader" commits to a strategy first, and a "follower" optimizes their response based on that commitment. SILQGames is a novel iterative algorithm designed to solve these games for systems with nonlinear dynamics and non-quadratic costs. It functions by repeatedly solving linear-quadratic approximations of the underlying game, allowing for efficient computation in complex physical scenarios.

Stackelberg Leadership Filter (SLF)

The SLF is an online Bayesian particle filter used to infer the latent leadership state. Each particle in the filter represents a hypothesis: either Agent A is the leader ($H_t = \mathcal{A}_1$) or Agent B is the leader ($H_t = \mathcal{A}_2$).

The Stackelberg Leadership Filter (SLF) process, showing the evolution of particles through prediction, measurement modeling via SILQGames, and resampling based on observed agent behavior.

At each time step, the filter simulates expected agent trajectories under both hypotheses using SILQGames. These predictions are compared against actual observations ($z_t$). Particles whose predictions align closely with observed behavior receive higher weights, enabling the filter to converge on the true leadership state even amidst noise and model mismatch.

Experimental Evaluation

The framework was evaluated across three simulated interactive environments:

  1. Shepherd and Sheep: A benchmark where a shepherd influences a sheep's position, tested in both linear and complex nonlinear configurations.
  2. Passing Scenario: A two-car interaction on a two-way road involving an overtake maneuver.
  3. Merging Scenario: A negotiation for right-of-way where two lanes merge into one.

Key Insights and Results

The core contribution of this work is the conceptualization of leadership as strategy influence. The leader is defined as the agent whose objective function is minimized subject to the follower's best response.

Dynamic Leadership Inference

The SLF successfully captures "leadership reversals" during interactive maneuvers. For example, in a passing scenario, the leadership probability shifts dynamically as the passing car initiates the move and the lead car reacts.

A passing maneuver visualization. The inset graph shows the real-time inference of leadership probability, demonstrating how the filter identifies shifts in influence as the cars interact.

Performance and Robustness

  • Convergence: SILQGames demonstrated reliable convergence in non-LQ scenarios, typically within 1100 iterations for complex cases.
  • Accuracy: The SLF correctly identified leaders with high confidence, provided a sufficient look-ahead horizon was used to distinguish intentional influence from random noise.
  • Computational Efficiency: A single SLF step with 50 particles took approximately 0.82 seconds on an AMD Ryzen 9 processor. While not yet meeting the <100ms latency required for high-speed highway driving, the parallelizable nature of particle filters suggests real-time performance is achievable.

Conclusion

This research provides a robust mathematical foundation for inferring leadership in nonlinear multi-agent interactions. By treating leadership as a structural element of a Stackelberg game, the framework moves beyond shallow heuristics toward a deeper understanding of agent intent. Future work will focus on parallelizing the SLF for real-time deployment, scaling to $N$-agent hierarchies, and integrating Inverse Reinforcement Learning to simultaneously infer agent objectives and leadership roles.

Summary updated

BibTeX
@article{khan2024leadership,
  author = {Hamzah Khan and David Fridovich-Keil},
  title = {{Leadership Inference for Multi-Agent Interactions}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2024},
  doi = {10.1109/lra.2024.3381469},
}
Figure