Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Attention Inference of Network Topology in Multi-Agent Systems

Akshay Kolli, Reza Azadeh, Kshitj Jerath

TL;DR

The results demonstrate that the presented data-driven graph attention machine learning model can identify the network topology in multi-agent systems, even when the underlying dynamic model is not known, as evidenced by the F1 scores achieved in the link prediction.

Abstract

Accurately identifying the underlying graph structures of multi-agent systems remains a difficult challenge. Our work introduces a novel machine learning-based solution that leverages the attention mechanism to predict future states of multi-agent systems by learning node representations. The graph structure is then inferred from the strength of the attention values. This approach is applied to both linear consensus dynamics and the non-linear dynamics of Kuramoto oscillators, resulting in implicit learning of the graph by learning good agent representations. Our results demonstrate that the presented data-driven graph attention machine learning model can identify the network topology in multi-agent systems, even when the underlying dynamic model is not known, as evidenced by the F1 scores achieved in the link prediction.

Graph Attention Inference of Network Topology in Multi-Agent Systems

TL;DR

The results demonstrate that the presented data-driven graph attention machine learning model can identify the network topology in multi-agent systems, even when the underlying dynamic model is not known, as evidenced by the F1 scores achieved in the link prediction.

Abstract

Accurately identifying the underlying graph structures of multi-agent systems remains a difficult challenge. Our work introduces a novel machine learning-based solution that leverages the attention mechanism to predict future states of multi-agent systems by learning node representations. The graph structure is then inferred from the strength of the attention values. This approach is applied to both linear consensus dynamics and the non-linear dynamics of Kuramoto oscillators, resulting in implicit learning of the graph by learning good agent representations. Our results demonstrate that the presented data-driven graph attention machine learning model can identify the network topology in multi-agent systems, even when the underlying dynamic model is not known, as evidenced by the F1 scores achieved in the link prediction.
Paper Structure (11 sections, 4 equations, 4 figures)

This paper contains 11 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Learning the Network Topology
  4. Model Architecture
  5. Model Training
  6. Results
  7. Network Topology Inference in Consensus Dynamics
  8. Network Topology Inference in Kuramoto Oscillators
  9. Learning of attention throughout training
  10. Concluding Remarks
  11. Future Work

Figures (4)

  • Figure 1: Examples of the simulation dynamics. (a) Consensus dynamics simulation (b) Kuramoto Oscillator simulation. The y-axis on both the plots represents the state information for each agent. The x-axis represents the time steps taken in the simulations. The plots on the bottom right represent the graph that the simulations operate on, with yellow signifying a connection in the adjacency matrix, and purple representing a zero.
  • Figure 2: The model consists of 4 key components (shown in blue) that get learned: The agent embeddings, the translation layer, the attention projection layer and the head.
  • Figure 3: (a), (c) Proposed method yields higher F1 scores (i.e., better performance) for graph inference with small number agents in both Consensus dynamics and Kuramoto oscillators. If data is limited, performance drops for systems with more agents. Performance calculated with 100 simulations worth of data. (b), (d) Using additional simulation data improves inference performance for larger systems as well.
  • Figure 4: Visualizing attention values through training stages, with the final column thresholding the attention value in the final epoch to give the predicted adjacency matrix. Reults demonstrate that most of the learning is done through the early epochs, with the later epochs adding to the final details.