⚡ electrical engineering

Maximum Likelihood Estimation for System Identification of Networks of Dynamical Systems

This paper proposes a consistent and efficient maximum likelihood estimation approach for identifying networks of dynamical systems that remains applicable even with partial node measurements and avoids the need for predictors to ensure computational efficiency.

Original authors: Anders Hansson, João Victor Galvão da Mata, Martin S. Andersen

Published 2026-02-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anders Hansson, João Victor Galvão da Mata, Martin S. Andersen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, bustling city where thousands of different machines (like traffic lights, power grids, or factory robots) are all talking to each other. Each machine has its own personality and rules for how it reacts to inputs, but they are all connected in a complex web.

The goal of this paper is to figure out exactly how each individual machine works just by listening to the conversations happening in the city. In the world of engineering, this is called System Identification.

Here is the breakdown of what the authors did, using simple analogies:

1. The Problem: The "Missing Microphone" Dilemma

Usually, to understand how a machine works, you need to record its input (what you tell it to do) and its output (what it does).

The Old Way: Previous methods required you to have a microphone on every single machine in the network. If you missed even one, the whole puzzle was unsolvable.
The New Way: The authors developed a method that works even if you only have microphones on some of the machines. You can still figure out how the silent machines are behaving, as long as the network follows certain logical rules (which they call "identifiability").

2. The Solution: The "Maximum Likelihood" Detective

The authors used a statistical strategy called Maximum Likelihood Estimation (MLE).

The Analogy: Imagine you are a detective trying to guess the rules of a game by watching people play. You don't know the rules, but you see the moves. You try to guess the rules that make the observed moves the most likely to happen.
The Innovation: Usually, detectives use a "predictor" (a crystal ball) to guess what happens next, then check if they were right. The problem is that in a complex network, building this crystal ball is mathematically messy and slow.
The Breakthrough: The authors found a way to solve the mystery without building the crystal ball first. They reformulated the math so they could look at the "missing data" (the unmeasured machines) directly. This is like solving a Sudoku puzzle by looking at the empty squares directly, rather than trying to guess every number one by one.

3. Why It Matters: Accuracy and Efficiency

The paper proves two big things about their new detective method:

Consistency: If you listen to the city for a long time, your guess about how the machines work will eventually become perfect. You won't be stuck with a "good enough" guess; you will get the true answer.
Efficiency: The method is as good as it possibly can be. It extracts the maximum amount of information from the data, meaning you need less time to get a reliable answer compared to other methods.

4. The "Partial View" Advantage

The authors tested their method on a network where they could only hear a few machines.

The Result: Even with missing information, their method successfully reconstructed the behavior of the whole network.
The Comparison: They compared their method to an older, faster method (called PEM).
- When they could hear everything, both methods worked well, but the older one was faster.
- When they could only hear some things, the older method failed or gave biased (wrong) answers. The new method, however, kept working and gave accurate results, even though it took a bit more computer power to do so.

Summary

Think of this paper as a new, smarter way to reverse-engineer a complex machine network.

Old Method: "I need to see every gear to know how the clock works."
New Method: "I can figure out how the hidden gears work just by watching the ones I can see, using a special mathematical trick that avoids getting stuck in complex calculations."

The authors show that this new trick is mathematically sound, guarantees the right answer with enough data, and works in situations where previous methods simply give up.

Technical Summary: Maximum Likelihood Estimation for System Identification of Networks of Dynamical Systems

Problem Statement
The paper addresses the challenge of identifying the open-loop transfer functions of networks of interconnected dynamical systems. While existing research in system identification of dynamic networks is categorized into topology estimation, single-module identification, and full network dynamics identification, this work focuses on the third category: identifying the full network dynamics for a given topology.

A primary limitation of existing Maximum Likelihood (ML) methods in this domain (e.g., [33]) is the requirement that all node signals in the network be measured. In many practical scenarios, measurements are only available for a subset of nodes. Furthermore, standard ML formulations for such partial observation problems often rely on predictors involving algebraic Riccati equations, leading to non-convex optimization problems that are difficult to solve numerically. The authors aim to establish an ML framework that is consistent and efficient even when measurements are unavailable for all network nodes, provided the network satisfies specific identifiability conditions.

Methodology
The authors propose a direct ML approach for identifying the parameters $(a, b, c, \lambda)$ of ARMAX models representing the individual systems within the network. The methodology proceeds through several key theoretical and algorithmic steps:

Model Formulation: The network is modeled as a collection of Single-Input Single-Output (SISO) ARMAX systems interconnected via a linear static feedback structure. The closed-loop system is described in state-space form, incorporating exogenous signals and noise.
Predictor and Innovation Form: The authors derive the innovation form of the closed-loop system, which defines the Kalman predictor. They establish conditions (Assumptions A2 and A3) under which the solution to the associated algebraic Riccati equation is unique and the predictor is stable.
Consistency and Efficiency: The paper proves that the ML estimates are consistent (converging to the true parameters as data length $N \to \infty$ ) and efficient (achieving the Cramér-Rao lower bound). This proof relies on the concept of "informativity" of the signals and "generic network identifiability." The authors demonstrate that while the open-loop transfer functions $(G)$ can be consistently identified under generic network identifiability, the noise dynamics $(H)$ generally cannot be identified unless specific structural conditions (such as full measurement of all nodes or specific network topologies) are met.
Equivalent Optimization Formulation: A significant methodological contribution is the reformulation of the ML problem. Standard formulations treat the problem as a constrained optimization involving the algebraic Riccati equation. The authors show that the problem can be equivalently formulated as an unconstrained optimization problem by treating the unmeasured states and noise sequences as latent variables (missing data). This approach, inspired by [15] and [32], eliminates the need for explicit predictor constraints or Riccati equation solvers during the optimization process, thereby facilitating more efficient numerical solutions.

Key Contributions
The paper summarizes its contributions as follows:

Theoretical Guarantees: It establishes the conditions for the consistency and efficiency of ML system identification for networks. This generalizes previous work on single systems to networks and provides alternative proofs for the ergodic case.
Unconstrained Optimization: It formulates the ML identification problem as an unconstrained optimization problem. This circumvents the numerical difficulties associated with the non-convex constraints typically found in predictor-based ML formulations.
Partial Measurement Capability: The method is applicable even when not all network nodes are measured, provided the network satisfies generic network identifiability. This contrasts with existing ML methods that require full node observability.

Results
The authors validate their approach using numerical experiments based on a network of dynamical systems (a specific topology from [16]).

Comparison with PEM: The proposed ML method is compared against Prediction Error Method (PEM) approaches. In cases where both methods are applicable (e.g., when specific signals allow for a linear predictor formulation), the ML method achieves fit values comparable to PEM.
Robustness to Partial Data: The ML method successfully identifies the network dynamics when only a subset of signals (e.g., $u_3$ ) is measured, a scenario where linear PEM formulations fail to identify all transfer functions without overfitting or bias.
Performance Metrics: While the ML method exhibits slightly higher computational time and a small rate of non-convergence (attributed to the non-convex nature of the problem) compared to PEM, it demonstrates lower Mean Squared Error (MSE) and bias in parameter estimates, particularly in scenarios with limited data where PEM tends to overfit.
Identifiability: The results confirm that the method can recover the open-loop transfer functions $G$ consistently, even when the noise dynamics $H$ cannot be fully recovered due to the lack of full network measurements.

Significance
The paper claims that its primary significance lies in providing a theoretically sound and numerically tractable ML framework for network identification that relaxes the strict requirement of full node observability. By reformulating the problem to avoid explicit predictor constraints, the authors enable the application of ML to a broader class of network identification problems. The work highlights that while linear PEM methods may be computationally faster, they can suffer from bias and overfitting in complex network scenarios or with short data sequences, whereas the proposed ML approach offers a more robust solution for achieving consistent and efficient estimates under general identifiability conditions. The authors note that their results extend trivially to cases where some transfer functions are known.

1. The Problem: The "Missing Microphone" Dilemma

2. The Solution: The "Maximum Likelihood" Detective

3. Why It Matters: Accuracy and Efficiency

4. The "Partial View" Advantage

Summary

More like this