Prediction with Approximated Gaussian Process Dynamical Models

Thomas Beckers,Sandra Hirche
DOI: https://doi.org/10.48550/arXiv.2006.14551
2021-11-30
Abstract:The modeling and simulation of dynamical systems is a necessary step for many control approaches. Using classical, parameter-based techniques for modeling of modern systems, e.g., soft robotics or human-robot interaction, is often challenging or even infeasible due to the complexity of the system dynamics. In contrast, data-driven approaches need only a minimum of prior knowledge and scale with the complexity of the system. In particular, Gaussian process dynamical models (GPDMs) provide very promising results for the modeling of complex dynamics. However, the control properties of these GP models are just sparsely researched, which leads to a "blackbox" treatment in modeling and control scenarios. In addition, the sampling of GPDMs for prediction purpose respecting their non-parametric nature results in non-Markovian dynamics making the theoretical analysis challenging. In this article, we present approximated GPDMs which are Markov and analyze their control theoretical properties. Among others, the approximated error is analyzed and conditions for boundedness of the trajectories are provided. The outcomes are illustrated with numerical examples that show the power of the approximated models while the the computational time is significantly reduced.
Systems and Control,Machine Learning
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the application challenges of Gaussian Process Dynamic Models (GPDMs) in control theory. Specifically, the paper focuses on how to deal with the non - Markovian characteristics of GPDMs and their impact on the properties of control theory. These problems mainly include: 1. **Theoretical problems**: - **Stability analysis**: Traditional control system analysis tools, such as Lyapunov stability analysis, are mainly applicable to systems with Markovian characteristics. Due to the non - Markovian characteristics of GPDMs, these tools are difficult to be directly applied to the stability analysis of GPDMs. - **Performance evaluation**: For non - Markovian systems, there are fewer existing performance evaluation tools, which makes the performance evaluation of control systems based on GPDMs difficult. 2. **Practical problems**: - **Computational complexity**: As the prediction time step increases, the number of past states and inputs that GPDMs need to consider keeps increasing, resulting in a rapid growth in the size of the covariance matrix. This not only increases the computation time but also brings about memory management problems, especially in long - time simulations. - **Memory problems**: Due to the need to store a large amount of past state and input data, long - time simulations may lead to out - of - memory problems. In order to overcome these challenges, the paper proposes an approximate GPDMs method, namely **Gaussian Process Approximate State - Space Model (GP - ASSM)**. By limiting the number of past states and inputs used for prediction, these models restore the Markovian characteristics, so that traditional control theory tools can be used for analysis and design while keeping the computation time constant. ### Main contributions 1. **Introducing the approximate GP - ASSM model**: By only considering a limited number of past states and inputs, the Markovian characteristics are restored. 2. **Analyzing the approximation error**: The upper bounds of the mean - square prediction error and the Kullback - Leibler divergence are provided to ensure that the prediction uncertainty of the approximate model does not underestimate the uncertainty of the real model. 3. **Guaranteeing control - related properties**: It is proved that the approximate model remains bounded in both open - loop and closed - loop cases. 4. **Case studies**: The behavior and advantages of the approximate model are demonstrated through two case studies, verifying its effectiveness and feasibility in practical applications. ### Mathematical representation - **Gaussian Process State - Space Model (GP - SSM)**: \[ x_{t + 1}=f(\xi_t)=\begin{cases} f_1(\xi_t)\sim\mathcal{GP}(m_1(\xi_t), k_1(\xi_t,\xi'_t)) \\ \vdots \\ f_{n_x}(\xi_t)\sim\mathcal{GP}(m_{n_x}(\xi_t), k_{n_x}(\xi_t,\xi'_t)) \end{cases} \] \[ y_t\sim p(y_t|x_t,\gamma_y) \] where \(\xi_t = [x_t^{\top}, u_t^{\top}]^{\top}\) is the combination of the state vector \(x_t\) and the input vector \(u_t\). - **Gaussian Process Nonlinear Output Error Model (GP - NOE)**: \[ y_{t + 1}=h(\zeta_t)=\begin{cases} h_1(\zeta_t)\sim\mathcal{GP}(m_1(\zeta_t), k_1(\zeta_t,\zeta'_t)) \\ \vdots \\ h_{n_y}(\zeta_t)\sim\mathcal{GP}(m_{n_y}(\zeta_t), k_{n_y}(\zeta_t,\zeta'_t)) \end{cases} \] where \(\zeta_t = [y_{t - n_{out}+1}^{\top},\ldots,y_t^{\top}, u_{t - n_{in}+1}^{\top},\ld