Learning Game-Theoretic Models of Multiagent Trajectories Using Implicit\n Layers

2020/08/17 by Philipp Geiger, Geiger, Philipp, Christoph-Nikolas Straehle +1 · 2 citations
Computer Science · #Time Series Analysis and Forecasting #Statistical and Computational Modeling #Bayesian Modeling and Causal Inference

paper · pdf · doi:10.48550/arxiv.2008.07303

Abstract

For prediction of interacting agents' trajectories, we propose an end-to-end\ntrainable architecture that hybridizes neural nets with game-theoretic\nreasoning, has interpretable intermediate representations, and transfers to\ndownstream decision making. It uses a net that reveals preferences from the\nagents' past joint trajectory, and a differentiable implicit layer that maps\nthese preferences to local Nash equilibria, forming the modes of the predicted\nfuture trajectory. Additionally, it learns an equilibrium refinement concept.\nFor tractability, we introduce a new class of continuous potential games and an\nequilibrium-separating partition of the action space. We provide theoretical\nresults for explicit gradients and soundness. In experiments, we evaluate our\napproach on two real-world data sets, where we predict highway driver merging\ntrajectories, and on a simple decision-making transfer task.\n

Cited by

• When the Correct Model Fails: The Optimality of Stackelberg Equilibria with Follower Intention Updates
• A taxonomy of strategic human interactions in traffic conflicts

Related

• Learning Two-agent Motion Planning Strategies from Generalized Nash Equilibrium for Model Predictive Control
• NashFormer: Leveraging Local Nash Equilibria for Semantically Diverse Trajectory Prediction
• Deep Interactive Motion Prediction and Planning: Playing Games with Motion Prediction Models
• Learning to Play Trajectory Games Against Opponents with Unknown Objectives
• Game Networks