Introduction
Modeling dynamical systems—those that evolve over time, such as robotic motion or environmental changes—is a cornerstone of AI and control theory. Traditionally, these systems are modeled using latent states that summarize historical data. A powerful alternative is the Predictive State Representation (PSR), where the state is defined by predictions of future observations rather than a summary of the past.
While PSRs work effectively for linear systems through Predictive Linear Gaussian (PLG) models, extending them to nonlinear, stochastic environments has proven difficult. Existing "Kernel PLG" methods often struggle with generalization and state-dependent noise. This paper introduces the Mixture of Predictive Linear Gaussian (MPLG) model to bridge this gap.
Methodology: The MPLG Model and Hybrid Inference
The authors propose two primary innovations to handle nonlinearity and stochasticity:
1. The MPLG Model
Instead of a single global linear model, the MPLG employs a weighted mixture of $J$ local PLGs. The weights are determined by Gaussian kernels centered in the state space. This architecture ensures that the model generalizes to a stable linear trend when encountering data far from the training set, unlike standard kernel methods that may decay to zero. Furthermore, each mixture component maintains its own noise statistics, allowing the model to capture uncertainty that varies with the system's state.
2. Hybrid Particle-Analytical Inference
Updating states in nonlinear systems typically requires computationally expensive approximations like Sigma-Point Approximations (SPAs). The authors introduce a "hybrid" approach: they use sigma-points only for the nonlinear state variables while performing exact analytical inference for the conditionally linear noise terms. This optimization significantly improves both speed and accuracy compared to naive sampling methods.
Experimental Results
The MPLG was evaluated against synthetic systems (e.g., Biped, Peanut, Spring) and real-world benchmarks (e.g., Santa Fe Laser, Mackey-Glass).
Key findings include:
- Performance: MPLG achieved top-tier performance (within 5% of the best MSE) on 4 out of 8 problems, particularly excelling in complex nonlinear dynamics like the "Biped" and "Peanut" tasks.
- Robustness: Qualitative analysis of error distributions demonstrates that MPLG is less sensitive to hyperparameter tuning than Kernel PLG, with errors consistently concentrated at low values.
- Memory: While Kernel Autoregressive (KAR) models performed well on purely autoregressive tasks (like Mackey-Glass), MPLG proved superior in scenarios where leveraging "state" provided an effectively infinite memory of the system's history.
Conclusion
The MPLG model successfully extends predictive state representations to nonlinear systems by combining local linear experts with an efficient hybrid inference scheme. By treating predictions as state, the model maintains a long-term memory of system dynamics while ensuring robust generalization through its mixture-of-experts structure. Future work will focus on automating model selection and scaling the framework to high-dimensional observation spaces.
Summary updated