When modeling complex dynamical systems from data, we often use state-space representations where a hidden internal state evolves over time based on inputs and its own previous values. Training these models on long data sequences presents a notorious challenge: small errors in predicting the next state compound rapidly over time, leading to exploding or vanishing gradients.
To circumvent this, a standard practice is to break long data sequences into shorter, truncated segments—a technique closely related to multiple shooting in control theory. However, this introduces a new dilemma: how do we determine the initial hidden state for each independent segment?
Traditionally, researchers treated the initial state of every single training segment as a free optimization parameter. While functional, this approach scales poorly with large datasets and completely fails when encountering unseen data, as there is no mechanism to infer the state of a new sequence without running a costly optimization loop.
The research direction detailed below solves this problem by introducing subspace encoders: neural networks trained to directly map a finite window of historical inputs and outputs to the system's latent state.
Learning the Reconstructability Map
The core reason subspace encoders work lies in a fundamental concept from control theory: observability. For a well-behaved dynamical system, the internal state at any given time is uniquely encoded within a sufficiently long window of its past inputs and outputs. The mathematical function that performs this inversion is known as a reconstructability map.
Instead of treating initial states as isolated variables, the authors use a feed-forward neural network to approximate this reconstructability map. As introduced by Gerben Beintema et al. in Nonlinear state-space identification using deep encoder networks, the encoder observes a local history window of past inputs and outputs to instantly estimate the starting latent state of a data segment.
Past History Window Latent State Future Predictions
[ u(t-P)...u(t), y(t-P)...y(t) ] ---> ( Subspace Encoder ) ---> x(t) ---> ( Predictor Network ) ---> y(t+1)...
By substituting free optimization parameters with a generalized neural mapping, this architecture enables highly efficient batch processing on modern hardware, drastically reducing training time while maintaining high simulation accuracy.
Stability in Continuous and Discrete Time
This framework, often referred to as the SUBNET (Subspace Encoder) approach, provides a robust foundation for identifying both discrete and continuous-time systems.
In Deep subspace encoders for nonlinear system identification, Gerben Izaak Beintema et al. demonstrate that combining a truncated prediction loss with a subspace encoder yields a highly stable and data-efficient training regime. Because the encoder is forced to learn a consistent state representation across different subsections of data, the model avoids the optimization drift typical of recurrent architectures.
This stability becomes even more pronounced when identifying continuous-time systems (such as Neural ODEs). As Gerben Izaak Beintema et al. show in Continuous-time identification of dynamic state-space models by deep subspace encoding, utilizing a neural encoder to initialize hidden states from a window of past observations provides a reliable starting point for numerical integrators. This approach mitigates the sensitivity to noise and initial conditions that frequently plagues continuous-time system identification.
To accelerate convergence, the encoder itself can be initialized with prior structural knowledge. For instance, Rishi Ramkannan et al. demonstrate in Initialization Approach for Nonlinear State-Space Identification via the Subspace Encoder Approach that extracting a linear reconstructability map from the system's Best Linear Approximation (BLA) provides an excellent weight initialization for the encoder network, particularly when dealing with weakly nonlinear systems.
From Training Trick to Real-Time State Observer
The utility of a subspace encoder extends far beyond accelerating model training; it fundamentally changes how models are deployed in downstream applications like Model Predictive Control (MPC).
In traditional control pipelines, tracking the non-physical, hidden states of a neural network requires designing complex, high-order recursive filters or state observers (such as Extended Kalman Filters). This introduces massive computational overhead.
Subspace encoders bypass this requirement entirely. Because the encoder is trained to map raw input-output histories directly to the internal state space, it functions as an instantaneous, zero-overhead state observer. As Jan H. Hoekstra et al. show in Computationally efficient predictive control based on ANN state-space models, leveraging this pre-trained encoder allows controllers to track latent states on the fly using only a sliding window of past measurements, enabling real-time, optimization-free state estimation during control loops.
Application Across Diverse Frameworks
The principle of direct state reconstruction via encoding has been successfully adapted across several distinct modeling paradigms:
- Physical and Geometric Constraints: The encoder-based training procedure integrates naturally with physics-informed architectures. Sarvin Moradi et al. combine subspace encoding with Port-Hamiltonian systems in Port-Hamiltonian neural networks with output error noise models, providing a scalable way to identify energy-preserving systems from noisy output measurements. Similarly, Jan H. Hoekstra et al. utilize past-window state reconstruction to optimize hybrid models structured as Linear Fractional Representations (LFRs) in Learning-based model augmentation with LFRs.
- Linear Parameter-Varying (LPV) Systems: In Deep-Learning-Based Identification of LPV Models for Nonlinear Systems, Chris Verhoek et al. deploy an encoder network to handle state initialization while simultaneously learning the scheduling maps that govern complex nonlinear transitions.
- Vehicle Dynamics: Real-world validation of this approach includes modeling highly nonlinear automotive systems. Gábor Rödönyi et al. demonstrate in Identification of the nonlinear steering dynamics of an autonomous vehicle that deep subspace encoders can successfully map historical driving data to internal steering states, outperforming traditional Gaussian Process and linear baselines.
Even outside classical system identification, the underlying philosophy—constraining latent states by forcing them to map to or from observable windows—recurs throughout sequence modeling. For example, the Predictive-State Decoders introduced by Arun Venkatraman et al. in Predictive-State Decoders: Encoding the Future into Recurrent Networks operate on a complementary duality, forcing RNN hidden states to directly decode into a sequence of future observations. In biological modeling, Finn Schmidt et al. apply an encoder network in Modeling dynamic neural activity by combining naturalistic video stimuli and stimulus-independent latent factors to infer latent brain states from subsets of neuronal responses, demonstrating the universal utility of using direct encoding networks to anchor hidden dynamic representations to observable data.
Key Foundational Work
Identification of the nonlinear steering dynamics of an autonomous vehicle (2021) — contributed 3 concept elements to this direction.
Nonlinear state-space identification using deep encoder networks (2021) — contributed 2 concept elements to this direction.
Predictive-State Decoders: Encoding the Future into Recurrent Networks (2017) — contributed 1 concept elements to this direction.