Date of Award

12-2017

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Dr. Ajay Gupta

Second Advisor

Dr. Cynthia J. Pietras

Third Advisor

Dr. Li Yang

Keywords

High performance computing, risky decision making, ensemble clustering, prediction in CUPS, toolkit

Abstract

The process or activity of making choices when subject to gain or loss can be understood as risky decision making (RDM). Risky Decisions consists of outcomes of decisions that may probabilistically result in unfavorable results. Every organism that lives faces this challenge and recent research suggests that there is a computational process involved in making these decisions. This has led to new approaches in the study of RDM. My dissertation is towards contributing to expand on the existing knowledge of RDM processes.

The core contribution of my work is an analysis and development of high performance computing techniques that improves contemporary research by providing new tools. An open source toolkit called RDMTk was developed as a part of this work that is available for use in the form of Software as a Service. RDMTk uses cloud based computing resources for running analysis code on demand by a researcher.

Computational challenges are addressed for a reinforcement learning algorithm currently used in identifying individual differences in RDM [71]. A central objective was an exhaustive analysis for improving this technology using shared and distributed memory. Algorithms for MPI (Message Passing Interface) based distributed memory and CUDA-GPU (Compute Unified Device Architecture-Graphic Processing Unit) based shared memory are developed and tested using extensive experiments. Our implementation on distributed architecture was able to achieve

almost a linear speedup (e.g. 44.79x using 48 MPI threads). And showed a 130x speedup for CPU-GPU based shared memory implementation over CPU only. We also discuss a novel Floor Tiles Planning theoretical approach to further reduce the computational overhead in RDM algorithms. This approach exploits spatial & temporal dependencies in computing resource allocation along with associated data dependencies. Data for our RDM research is collected through open source RDMTk toolkit, developed as a part of the dissertation work.

Access Setting

Dissertation-Open Access

Download

Share

COinS