Abstract
This paper proposes a general switching dynamical system model, and a custom majorization-minimization-based algorithm EM++ for identifying its parameters. For certain families of distributions, such as Gaussian distributions, this algorithm reduces to the well-known expectation-maximization method. We prove global convergence of the algorithm under suitable assumptions, thus addressing an important open issue in the switching system identification literature. The effectiveness of both the proposed model and algorithm is validated through extensive numerical experiments.
Introduction
Many real-world systems exhibit "switching" behaviors, where the underlying dynamics change between different modes. Examples include robotic locomotion across varying terrains or chemical processes undergoing phase transitions. Identifying the parameters of these switching dynamical systems from observed data is challenging because the switching times and active modes are typically latent variables. While the Expectation-Maximization (EM) algorithm is a standard tool for latent variable models, it often lacks global convergence guarantees for switching systems and is frequently restricted to Gaussian noise assumptions.
The EM++ Framework
To address these limitations, we introduce EM++, a robust parameter learning framework based on the Majorization-Minimization (MM) principle. EM++ generalizes the classic EM algorithm by constructing a convex surrogate function at each iteration that upper-bounds the non-convex negative log-likelihood.
A key strength of EM++ is its flexibility across various probability distributions. While it recovers standard EM for Gaussian noise, it extends naturally to the broader exponential family and heavy-tailed distributions such as Student’s t and Laplace. This versatility is summarized in the following table:
To maintain computational efficiency, EM++ employs a forward-backward algorithm to evaluate the surrogate function in linear time relative to the data length, avoiding the exponential complexity typically associated with switching sequences. Furthermore, we provide a theoretical guarantee of convergence to a stationary point under the Kurdyka-Lojasiewicz (KL) condition.
Experimental Validation
The EM++ framework was evaluated against synthetic and real-world benchmarks, including switched Markov ARX systems and a cart-and-bumper model.
Robustness and Efficiency
EM++ demonstrates significant robustness to outliers by utilizing heavy-tailed distributions. In experiments with 5% outliers, a Student’s t-based EM++ model maintained high accuracy where Gaussian-based models failed. In terms of optimization efficiency, EM++ reached convergence in significantly fewer iterations compared to standard gradient-based optimizers like Adam or BFGS.
Real-World Performance
On the "Coupled Electric Drives" benchmark, EM++ outperformed Neural Network ARX and Gaussian Process ARX models in terms of Root Mean Square Error (RMSE). Its ability to learn parametric switching mechanisms also makes it suitable for real-time control applications, as it avoids the need for per-prediction optimization.
Discussion and Future Work
While EM++ provides a powerful and theoretically grounded framework for switching system identification, several avenues for future research remain. The question of "identifiability"—whether the recovered parameters are unique—remains an open theoretical challenge. Additionally, extending the framework to handle continuous latent variables and developing an "online" version for real-time parameter updates would further enhance its utility in adaptive control systems.
Summary updated