Learning Game-Theoretic Models of Multiagent Trajectories Using Implicit Layers

Philipp Geiger, Christoph-Nikolas StraehleAAAI 2021 · DOI

Introduction

Predicting human interaction in traffic—such as the complex "who goes first" negotiation during highway merging—remains a significant hurdle for autonomous vehicle (AV) development. While purely data-driven neural networks offer flexibility, they often lack interpretability and formal safety guarantees. Conversely, classical game theory provides a rigorous mathematical framework but struggles to scale to the continuous, high-dimensional movements of real-world driving.

In their AAAI-21 paper, "Learning Game-Theoretic Models of Multiagent Trajectories Using Implicit Layers," Philipp Geiger and Christoph-Nikolas Straehle from the Bosch Center for AI introduce the Trajectory Game Learner (TGL). This hybrid architecture embeds game-theoretic reasoning directly into a neural network using an implicit layer, bridging the gap between flexible learning and structured reasoning.

The Challenge of Multiagent Interaction

Most trajectory prediction models treat agents as independent or rely on "black-box" social pooling. These approaches often fail to capture multi-modality—the reality that a single situation can have multiple valid outcomes, such as Car A merging before or after Car B. Furthermore, when a prediction fails, black-box models offer little insight into whether the error stemmed from a misunderstanding of intent or a failure to reason about the interaction itself.

Solving games with continuous actions is computationally intensive and typically does not provide the gradients necessary for end-to-end neural network training. TGL addresses this by framing the problem as a differentiable optimization task.

The TGL Architecture

The core innovation of TGL is the Game Solver Implicit Layer. Unlike standard layers with fixed functions (like ReLU), this layer defines its output as the solution to an equilibrium problem.

The TGL architecture pipeline, from past trajectories to predicted joint modes.

The model operates through three primary components:

  1. Preference Revelation Net: An LSTM or MLP processes past trajectories to "reveal" latent game parameters, such as an agent's relative weighting of speed versus safety.
  2. Equilibrium Refinement Net: Because games can have multiple equilibria, this network predicts the likelihood of different "modes" (e.g., "Red car yields").
  3. Implicit Layer: By applying the Implicit Function Theorem, the model differentiates through the game's optimality conditions. This allows the network to backpropagate errors from the final coordinates all the way to the preference parameters.

To ensure mathematical tractability, the authors frame interactions as Common-Coupled Potential Games. In these games, agents share certain interests (like collision avoidance) while maintaining individual preferences. This structure ensures that local Nash Equilibria (NE) correspond to the local optima of a single "potential function," making them significantly easier to solve.

Datasets and Evaluation

The researchers evaluated TGL on two highway merging benchmarks:

  • highD: A standard dataset of vehicle trajectories recorded via drones on German highways.
  • HEE (Highway Eagle Eye): A new dataset introduced in this work, featuring ~12,000 trajectories (~4 hours of footage) specifically at highway on-ramps.

Key Insights and Results

A major finding is that local Nash Equilibria naturally correspond to intuitive behavioral modes. For instance, in a merge scenario, one local NE might represent an agent accelerating to lead, while another represents braking to follow.

Visualization of two distinct local Nash Equilibria (modes) for a merging scenario.

Performance Metrics

On the highD dataset, the TGL-D variant demonstrated superior accuracy over a 7-second horizon:

  • TGL-D: 2.9m Mean Absolute Error (MAE)
  • CS-LSTM: 5.0m MAE
  • MFP: 5.2m MAE

Beyond accuracy, TGL offers strategic transferability. Because the model learns an underlying utility function, it can be integrated into a robotic decision-making pipeline. An AV can use the learned "game" to select actions that maximize its own utility while anticipating how human agents will react based on their learned preferences.

Conclusion

TGL represents a significant step toward interpretable, multi-modal trajectory prediction. By treating traffic as a potential game and leveraging implicit layers for differentiability, the authors provide a framework that "understands" the rules of the road. Future work aims to relax concavity assumptions and explore Markov Games to better account for temporal uncertainty.

Summary updated

BibTeX
@inproceedings{geiger2021learning,
  author = {Philipp Geiger and Christoph-Nikolas Straehle},
  title = {{Learning Game-Theoretic Models of Multiagent Trajectories Using Implicit Layers}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i6.16628},
}
Figure