Introduction
Predicting the future states of dynamical systems—such as maritime vessels or industrial robotic arms—is a fundamental challenge in automation and control. While deep learning models like Long Short-Term Memory (LSTM) networks have set new benchmarks for accuracy, they are often criticized as "black boxes" that lack formal stability guarantees and interpretability. Conversely, traditional control-theoretic models like switched linear systems offer mathematical elegance and stability but frequently fail to capture the complex nonlinearities of real-world data.
This paper introduces ReLiNet (Recurrent Linear Parameter-Varying Network), a novel architecture designed to bridge this gap. ReLiNet provides high-accuracy multistep predictions that are formally stable and inherently explainable by design.
Methodology
ReLiNet employs a "master-slave" architecture to combine the representational power of recurrent neural networks (RNNs) with the transparency of linear models.
The Master-Slave Architecture
The "Master" component is an RNN that processes control inputs to generate time-varying parameters (matrices) for a "Slave" linear model. The Slave model then uses these parameters to predict the next state based on the current state and control inputs. Because the final prediction is the result of a linear operation, the model is explainable by default; the contribution of each input to the output can be derived mathematically without relying on post-hoc approximation tools like LIME or SHAP.
Formal Stability with StableReLiNet
To address the risk of error accumulation in multistep predictions, the authors propose StableReLiNet. By enforcing specific mathematical constraints—ensuring system matrices are simultaneously diagonalizable and bounding their eigenvalues via a $tanh$ activation—the model is proven to be globally exponentially stable. This ensures that the model's outputs remain bounded and converge toward a steady state under constant inputs, a critical requirement for safety-critical applications.
Experimental Evaluation
The models were evaluated on two complex benchmarks: a Ship Maneuvering simulation (including an out-of-distribution set) and a real-world Industrial Robot joint angle dataset.
Accuracy and Generalization
ReLiNet achieves Normalized Root Mean Squared Error (NRMSE) scores comparable to state-of-the-art LSTMs. Notably, in the Ship OOD task, ReLiNet outperformed the LSTM across all prediction horizons, suggesting that its underlying linear structure provides superior generalization to unseen conditions compared to standard recurrent architectures.
Explainability and Faithfulness
A key advantage of ReLiNet is its "faithfulness." In explainable AI, infidelity measures how much an explanation deviates from the model's actual logic. ReLiNet achieves an infidelity score of zero, meaning its explanations are perfectly faithful representations of its internal mechanics. In contrast, post-hoc methods like LIME applied to LSTMs showed significantly higher infidelity, particularly on the ship datasets.
Discussion and Conclusion
The results demonstrate that ReLiNet successfully balances performance and safety. While StableReLiNet incurs a measurable "safety tax" in terms of accuracy compared to its unconstrained counterpart, it remains significantly more accurate than traditional linear baselines while providing rigorous stability guarantees. ReLiNet represents a significant advancement toward trustworthy AI in physical systems, proving that transparency and predictive power are not mutually exclusive.
Summary updated