Efficient Modeling of Multi-Agent Systems via Structure-Aware Physics-Informed Learning
Organizations: Division of Decision and Control Systems, KTH Royal Institute of Technology, Stockholm, Sweden · Department of Electronic & Electrical Engineering, University College London, United Kingdom
Abstract
Modeling large multi-agent systems poses challenges for trajectory inference and parameter identification, as network size and coupling increase computational demands and partial observations require reconstructing unobserved states. We propose MAS-PINN, a neural-network architecture with a tailored training algorithm for a physics-informed learning framework that leverages the network topology and the agents' similarity to reduce computational and memory costs of simulating and identifying multi-agent systems. Its architecture represents agent trajectories through shared temporal features and agent-specific coefficients. Training alternates between learning features, fitting coupled trajectories, and estimating physical parameters. A matrix-free update exploits local interactions to fit trajectories without forming large global matrices. Experiments on Lorenz rings, Lotka--Volterra networks, and Rössler oscillators show improved trajectory and parameter accuracy over baselines, including inference in large systems and identification under partial observations or unknown topology, alongside substantial computational savings in coefficient fitting.
Figures & tables
| Method | Trajectory | Final state | Trajectory | Final state |
|---|---|---|---|---|
| Vanilla PINN | ||||
| PirateNets | ||||
| MAS-PINN | ||||
| Coefficient solver | LM (s) | Total (s) | Peak memory (MB) |
|---|---|---|---|
| Matrix-free, 30 CG iterations | 1.20 | 1.78 | 32 |
| Matrix-free, 60 CG iterations | 2.25 | 2.82 | 32 |
| Matrix-free, 100 CG iterations | 3.64 | 4.26 | 32 |
| Dense LM, Cholesky | 77.77 | 78.44 | 3194 |
| Noise | Method | Trajectory | AUC | ||
|---|---|---|---|---|---|
| MAS-PINN | |||||
| HyperPINN | |||||
| MHE | |||||
| MAS-PINN | |||||
| HyperPINN | |||||
| MHE |
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | Value |
|---|---|
| Updates per outer iteration | 1 Adam step on , 1 LM step on , and 1 parameter solve for identification only |
| Initial damping multiplier | |
| LM damping | |
| Accepted-step multiplier | |
| Rejected-step multiplier | |
| Acceptance threshold | Actual decrease predicted decrease |
| Setting | MAS-PINN | Vanilla PINN | PirateNets |
|---|---|---|---|
| Time interval | |||
| Sites | |||
| State dimension, forcing | , | ||
| Initial state | Spin-up of time units from , | ||
| Collocation points | |||
| Network | MLP | MLP | PirateNet |
| Setting | MAS-PINN | HyperPINN |
|---|---|---|
| Time interval | ||
| Sites, species, diffusion | , , | |
| Interaction matrix | ||
| Graph | Ring on the sites plus the edges | |
| Dispersal weights | ||
| Observed sites | ||
| Setting | MAS-PINN | HyperPINN |
|---|---|---|
| Time interval | ||
| Oscillators, local parameters | , , | |
| Pairwise interactions | , strength | |
| Three-body interactions | , strength | |
| Candidate hyperedges | All pairs and triples | |
| Observations | ||
| Matrix-free CG | Dense LM | |||||
| Factor | Value | Op. (ms) | LM (s) | Mem. (MB) | LM (s) | Mem. (MB) |
| 4 | 0.35 | 3.63 | 2.8 | 10.9 | 828 | |
| 8 | 0.33 | 3.64 | 5.7 | 77.8 | 3169 | |
| 16 | 0.45 | 5.06 | 11.6 | 594.0 | 12387 | |
| 32 | 0.58 | 6.38 | 23.6 | — | — | |
| 64 | 1.36 | 14.39 | 59.5 | — | — | |
Multi-Agent Systems
Multi-Agent Systems studies how autonomous agents coordinate, compete, communicate, and learn in shared or interdependent tasks.
See all 727 papers