The past does matter: correlation of subsequent states in trajectory predictions of Gaussian Process models

Steffen Ridderbusch, Sina Ober-Blöbaum, Paul GoulartUAI 2023

Presented at UAI 2023, this research identifies and resolves a fundamental flaw in how we predict the long-term behavior of dynamical systems modeled by Gaussian Processes (GPs). By accounting for hidden correlations between the system state and the model, the authors introduce a novel solver that provides accurate uncertainty estimates where traditional methods fail.

The Problem: The Independence Assumption and Variance Collapse

When using a Gaussian Process to model a dynamical system—such as a robot's trajectory or a chemical reaction—we are learning a distribution over possible functions. This represents epistemic uncertainty: our lack of knowledge due to limited data. To predict a trajectory, we typically start at an initial state and iteratively apply the GP model to find subsequent states.

The core challenge lies in "uncertainty propagation." As we move forward in time, the uncertainty about our current position must be combined with the uncertainty of the GP model itself. Historically, researchers have used "moment matching" to approximate these distributions. However, these methods make a critical error: they assume that the current state $X_n$ and the GP function $f$ are independent at each step.

In reality, $X_n$ was reached by already using $f$ in previous steps. Ignoring this correlation leads to a phenomenon called variance collapse. Counter-intuitively, in standard GP trajectory prediction, smaller time steps actually make the variance error worse. This contradicts basic numerical analysis principles, where smaller steps usually yield higher accuracy.

Variance collapse in standard GP trajectory prediction: as the integration step size $h$ decreases, the predicted variance incorrectly shrinks, diverging from the ground truth sampling.

PULL: Propagating Uncertainty through Local Linearization

To fix this, the authors propose PULL, a new class of solvers designed to track these vital correlations through time. The method rests on three pillars:

  1. Piecewise Linearization: Instead of treating the GP as a black-box nonlinear map, PULL approximates the GP as a series of local linear models at each time step.
  2. Accounting for the Past: Unlike previous methods that "restart" the uncertainty at every step, PULL uses a "telescope sum" approach to track the cross-covariance between the current state and the model parameters. This captures the "missing link" of how the past affects the present.
  3. Exact Propagation: Because the local model is linear and the GP is Gaussian, the uncertainty can be propagated exactly through the linearized step without needing expensive Monte Carlo sampling.

Key Results and Performance

PULL is the first approximate GP trajectory method where the error consistently decreases as the integration step size $h$ gets smaller. In nonlinear tests—such as a system defined by $x' = x \cos(x)$—PULL’s variance predictions closely match ground truth results obtained from thousands of Monte Carlo samples, whereas standard methods significantly underestimate the risk.

PULL accurately predicts the trajectory distribution of a nonlinear system, matching the mean and standard deviation of expensive Monte Carlo samples.

Beyond accuracy, PULL is highly efficient. It provides sampling-quality results in milliseconds (roughly 14ms on a standard CPU), while high-quality Monte Carlo sampling can take seconds or minutes. This makes it a viable candidate for real-time applications in robotics and control.

Limitations and Future Outlook

While PULL represents a significant step forward, it currently assumes the state distribution remains Gaussian (unimodal). If a trajectory reaches a "fork in the road" (an unstable fixed point), PULL will average the paths into a single Gaussian rather than capturing a bimodal distribution. Future work will focus on extending PULL to higher-order solvers like Runge-Kutta and scaling the cross-covariance tracking to high-dimensional state spaces, such as multi-DOF drones or complex robotic manipulators.

Summary updated

BibTeX
@inproceedings{ridderbusch2023past,
  author = {Steffen Ridderbusch and Sina Ober-Blöbaum and Paul Goulart},
  title = {{The past does matter: correlation of subsequent states in trajectory predictions of Gaussian Process models}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2023},
}
Figure