Learning Optimal Interaction Weights in Multi-Agents Systems

TL;DR

This paper introduces a framework for learning interaction weights in multi-agent systems using inverse optimal control, capturing complex behaviors from trajectory data and analyzing how network topology influences coordination.

Contribution

It develops a novel graph-based inverse optimal control method for multi-agent interactions, incorporating spatial-temporal dynamics and providing conditions for optimality.

Findings

Successfully recovers interaction weights from trajectory data

Demonstrates effectiveness in multi-agent formation control

Analyzes the impact of network topology on coordination

Abstract

This paper presents a spatio-temporal inverse optimal control framework for understanding interactions in multi-agent systems (MAS). We employ a graph representation approach and model the dynamics of interactions between agents as state-dependent edge weights in a consensus algorithm, incorporating both spatial and temporal dynamics. Our method learns these edge weights from trajectory observations, such as provided by expert demonstrations, which allows us to capture the complexity of nonlinear, distributed interaction behaviors. We derive necessary and sufficient conditions for the optimality of these interaction weights, explaining how the network topology affects MAS coordination. The proposed method is demonstrated on a multi-agent formation control problem, where we show its effectiveness in recovering the interaction weights and coordination patterns from sample trajectory data.