One of the most important method of handling uncertain knowledge is Bayesian network. It is successfully applied in many areas in life including bioinformatics, financial systems and warfare. The main strength of Bayesian statistic is its ability to efficiently represent multivariate distributions through assumption of independence between variables. In additional in learning it can use a priori expert knowledge with statistical data. At the same time the model identification complexity is super-exponential, to overcome this issue we use Monte Carlo methods. Our aim is to explore structural features and approximate their a posteriori probabilities.
This thesis discusses a highly parallelized implementation of ordering MCMC (Markov Chain Monte Carlo) using OpenCL heterogeneous programing language and MPI (Message Passing Interface). While it is mainly designed for high-throughput computing systems where low network latency is granted it performs great for desktop simulation in case of small variables.
The features of the program will be presented using genomic data.