ACCELERATED LEARNING OF GAUSSIAN PROCESS MODELS Bojan Musizza 1 , Dejan Petelin 1 , Juˇ s Kocijan 1,2 1 Joˇ zef Stefan Institute Jamova 39, Ljubljana, Slovenia 2 University of Nova Gorica Vipavska 13, Nova Gorica, Slovenia dejan.petelin@ijs.si(Dejan Petelin) Abstract The Gaussian process model is an example of a flexible, probabilistic, nonpara- metric model with uncertainty predictions. It offers a range of advantages for mod- elling from data and has been therefore used also for dynamic systems identifica- tion. One of the noticeable drawbacks of the system identification with Gaussian process models is computation time necessary for modelling. The modelling pro- cedure involves the inverse of covariance matrix which is as large as the length of input samples vector. The computation time for this inverse regardless of the use of efficient algorithm is rising with the third power of input data number. Inten- sive research is going on for finding algorithms that would accelerate the training of Gaussian process models. The purpose of this paper is to show approach from the used hardware point of view. The assessment of computational efficiency of two different hardware platforms for GP model identification are given in the pa- per. These are: single core personal computer and personal computer with graphic card used for computations. The assessment has been done with comparison of computational load on a toy case study of nonlinear dynamic system identifica- tion. The assessment reveals that the parallel computation solutions are efficient for larger amount of data when the initial and communication overhead of parallel computation becomes sufficiently small part of the whole process. Keywords: Gaussian process models, Dynamic system models, System identification, Sim- ulation. Presenting Author’s Biography Dejan Petelin received the M.Sc. degree in computer science and infor- matics from the Faculty of Computer Science and Informatics, University of Ljubljana. He is currently a Ph.D. student at the Department of Sys- tems and Control, Jozef Stefan Institute in Ljubljana. His main research interests are machine learning methods and their application for dynamic systems modelling.