|
|
Overcomplete Mathematical Models with Applications
Tonner, Jaromír ; Witkovský,, Viktor (oponent) ; Martišek, Dalibor (oponent) ; Rajmic, Pavel (oponent) ; Veselý, Vítězslav (vedoucí práce)
Chen, Donoho a Saunders (1998) deal with the problem of sparse representation of vectors (signals) by using special overcomplete (redundant) systems of vectors spanning this space. Typically such systems (also called frames) are obtained either by refining existing basis or merging several such bases (refined or not) of various kinds (so-called packets). In contrast to vectors which belong to a finite-dimensional space, the problem of sparse representation may be formulated within a more general framework of (even infinite-dimensional) separable Hilbert space (Veselý, 2002b; Christensen, 2003). Such functional approach allows us to get more precise representation of objects from such space which, unlike vectors, are not discrete by their nature. In this Thesis, I attack the problem of sparse representation from overcomplete time series models using expansions in the Hilbert space of random variables of finite variance. A numerical study demonstrates benefits and limits of this approach when applied to generalized linear models or to overcomplete VARMA models of multivariate stationary time series, respectively. After having accomplished and analyzed a lot of numerical simulations as well as real data models, we can conclude that the sparse method reliably identifies nearly zero parameters allowing us to reduce the originally badly conditioned overparametrized model. Thus it significantly reduces the number of estimated parameters. Consequently there is no care about model orders the fixing of which is a common preliminary step used by standard techniques. For short time series paths (100 or less samples), the sparse parameter estimates provide more precise predictions compared with those based on standard maximum likelihood estimators from MATLAB's System Identification Toolbox (IDENT). For longer paths (500 or more), both techniques yield nearly equal prediction paths. On the other hand, solution of such problems requires more sophistication and that is why a computational speed is larger, but still comfortable.
|
|
|
Overcomplete Mathematical Models with Applications
Tonner, Jaromír ; Witkovský,, Viktor (oponent) ; Martišek, Dalibor (oponent) ; Rajmic, Pavel (oponent) ; Veselý, Vítězslav (vedoucí práce)
Chen, Donoho a Saunders (1998) deal with the problem of sparse representation of vectors (signals) by using special overcomplete (redundant) systems of vectors spanning this space. Typically such systems (also called frames) are obtained either by refining existing basis or merging several such bases (refined or not) of various kinds (so-called packets). In contrast to vectors which belong to a finite-dimensional space, the problem of sparse representation may be formulated within a more general framework of (even infinite-dimensional) separable Hilbert space (Veselý, 2002b; Christensen, 2003). Such functional approach allows us to get more precise representation of objects from such space which, unlike vectors, are not discrete by their nature. In this Thesis, I attack the problem of sparse representation from overcomplete time series models using expansions in the Hilbert space of random variables of finite variance. A numerical study demonstrates benefits and limits of this approach when applied to generalized linear models or to overcomplete VARMA models of multivariate stationary time series, respectively. After having accomplished and analyzed a lot of numerical simulations as well as real data models, we can conclude that the sparse method reliably identifies nearly zero parameters allowing us to reduce the originally badly conditioned overparametrized model. Thus it significantly reduces the number of estimated parameters. Consequently there is no care about model orders the fixing of which is a common preliminary step used by standard techniques. For short time series paths (100 or less samples), the sparse parameter estimates provide more precise predictions compared with those based on standard maximum likelihood estimators from MATLAB's System Identification Toolbox (IDENT). For longer paths (500 or more), both techniques yield nearly equal prediction paths. On the other hand, solution of such problems requires more sophistication and that is why a computational speed is larger, but still comfortable.
|
host ::
RSS.