Note: This content is accessible to all versions of every browser. However, this browser does not seem to support current Web standards, preventing the display of our site's design details.


Controller learning for structured Gaussian-Process Models


N. Hail

Master Thesis, SS 16 (10524)

Gaussian Process (GP) Models have been proposed as a data-driven, parameter-free approach for system identification. Based on these type of models, algorithms like PILCO (Probabilistic Inference for Learning Control) have been derived, which optimize controller parameters for a given GP model and a given controller structure. Although no theoretical guarantees for convergence exist to this point, comparable or superior performance in comparison to traditional reinforcement learning techniques has been reported for certain applications.
However, training GP models for high-dimensional systems is prohibitive so far, due to the large number of training data necessary to represent the dynamics in a high-dimensional space. We propose a new framework for structured GP models, primarily motivated by applications in traffic modeling and control. For traffic models, large parts of the systems dynamics are perfectly known, but certain parts of the dynamics, between small subsets of states, are affected by heavy uncertainty. In our framework, we allow to define systems as a combination of linear dynamics (assumed to be known a-priori) and unknown, preferably low-dimensional GPs. We derive and implement the equations necessary to train controller parameters for such models in a PILCO-like setting and demonstrate the efficiency of our approach in simulations.

Supervisors: Marius Schmitt, Chithrupa Ramesh, Prof. John Lygeros


Type of Publication:

(12)Diploma/Master Thesis

File Download:

Request a copy of this publication.
(Uses JavaScript)
% Autogenerated BibTeX entry
@PhdThesis { Xxx:2016:IFA_5498
Permanent link