Architectures and GPU-Based Parallelization for Online Bayesian Computational Statistics and Dynamic Modeling
Recent work demonstrates that coupling Bayesian computational statistics methods with dynamic models can facilitate the analysis of complex systems associated with diverse time series, including those involving social and behavioural dynamics. Particle Markov Chain Monte Carlo (PMCMC) methods constitute a particularly powerful class of Bayesian methods combining aspects of batch Markov Chain Monte Carlo (MCMC) and the sequential Monte Carlo method of Particle Filtering (PF). PMCMC can flexibly combine theory-capturing dynamic models with diverse empirical data. Online machine learning is a subcategory of machine learning algorithms characterized by sequential, incremental execution as new data arrives, which can give updated results and predictions with growing sequences of available incoming data. While many machine learning and statistical methods are adapted to online algorithms, PMCMC is one example of the many methods whose compatibility with and adaption to online learning remains unclear. In this thesis, I proposed a data-streaming solution supporting PF and PMCMC methods with dynamic epidemiological models and demonstrated several successful applications. By constructing an automated, easy-to-use streaming system, analytic applications and simulation models gain access to arriving real-time data to shorten the time gap between data and resulting model-supported insight. The well-defined architecture design emerging from the thesis would substantially expand traditional simulation models' potential by allowing such models to be offered as continually updated services. Contingent on sufficiently fast execution time, simulation models within this framework can consume the incoming empirical data in real-time and generate informative predictions on an ongoing basis as new data points arrive. In a second line of work, I investigated the platform's flexibility and capability by extending this system to support the use of a powerful class of PMCMC algorithms with dynamic models while ameliorating such algorithms' traditionally stiff performance limitations. Specifically, this work designed and implemented a GPU-enabled parallel version of a PMCMC method with dynamic simulation models. The resulting codebase readily has enabled researchers to adapt their models to the state-of-art statistical inference methods, and ensure that the computation-heavy PMCMC method can perform significant sampling between the successive arrival of each new data point. Investigating this method's impact with several realistic PMCMC application examples showed that GPU-based acceleration allows for up to 160x speedup compared to a corresponding CPU-based version not exploiting parallelism. The GPU accelerated PMCMC and the streaming processing system can complement each other, jointly providing researchers with a powerful toolset to greatly accelerate learning and securing additional insight from the high-velocity data increasingly prevalent within social and behavioural spheres. The design philosophy applied supported a platform with broad generalizability and potential for ready future extensions. The thesis discusses common barriers and difficulties in designing and implementing such systems and offers solutions to solve or mitigate them.
Particle Filterting, Particle MCMC, MCMC, GPU
Master of Science (M.Sc.)