Abstract
dc:descriptionIn this thesis we use a Bayesian nonparametric prior with simple weights, namely the Geometric Stick--Breaking (GSB) random probability measure to deal with the problem of reconstruction and prediction of stochastic discretized nonlinear dynamical systems.In the first half of the thesis we propose a Bayesian nonparametric mixture model for the reconstruction and prediction from observed time series data, of discretized stochastic dynamical systems, based on Markov Chain Monte Carlo (MCMC) methods. Our approach is nonparametric in the sense that we model the noise component with a highly flexible family of density functions. While the common assumption is the normality of the noise process, here we model the noise component as an infinite mixture of Normal kernels with the mixing weights driven by a random probability measure sampled from a GSB process. In the second half we present a new approach on the joint estimation of partially exchangeable observations by constructing pairwise dependence between a finite collection of random density functions, each of which is modeled as a mixture of GSB processes. This approach is based on a new random central masses version of the Pairwise Dependent Dirichlet Process prior mixture model. We show that modelling with Pairwise Dependent Geometric Stick--Breaking Processes (PDGSBP) is sufficient for estimation and prediction purposes.We also propose a Bayesian nonparametric mixture model for the full reconstruction of a finite collection of dynamical equations, given observed dynamically--noisy--corrupted chaotic time series based on PDGSBP mixture priors. Under the assumption that the each set of dynamical equations has a deterministic part with known functional form and that the noise processes are independent and identically distributed from some unknown zero mean process which may have common characteristics, we jointly estimate the parameters of the dynamical systems and perform density estimation of noise components. We show that if there is at least one sufficiently large data set, using borrowing--of--strength prior specifications we are able to reconstruct those dynamical processes that are responsible for the generation of time series with small sample sizes which are inadequate for an independent reconstruction. Our contention is that modeling with GSB random probability measures is sufficient for estimation and prediction purposes. The proposed MCMC algorithms are faster and easier to implement than their Dirichlet process based counterparts. The advantages of the use of such a simple random probability measure in Bayesian nonparametric inference in terms of sufficiency and time complexity are illustrated in both synthetic and real data sets.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Merkatas, Christos
- Μέρκατας, Χρίστος
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/43751