high performance computing on graphics processing units: hgpu.org

The applicability and effectiveness of adaptive design optimization (ADO) in selecting optimal stimuli or designs for experimental trials has been well demonstrated in several content areas of cognitive psychology (Myung & Pitt, 2009; Cavagnaro et al, 2010). On the other hand, when applying ADO to real-time, online experiments such as psychophysical experiments with human subjects, the speed and efficiency of ADO algorithms remain a challenge. Significant computational costs associated with the implementation of ADO have been a major obstacle in applications with time constraints. This study explores the acceleration of ADO computations by taking advantage of increasingly powerful hardware. Given the recent interest in and use of parallel computing offered by graphics processing units (GPUs), the first objective of this study is to examine the feasibility and performance of GPU-based ADO in the estimation of the threshold and slope of the visual psychometrical model in Kontsevich and Tyler (1999), and compare the runtime to that of the standard, CPU-based ADO computation. To this end, a GPU-based grid search algorithm was implemented and compared with the identical algorithm run on a single CPU chip. The results showed that GPU computing provided a dramatic reduction in runtime, which substantiates the feasibility of using GPU computing to improve the computational speed of ADO algorithms. A second objective of the study examines the efficiency of different ADO algorithms, or their ability to identify optimal designs, reflected by the number of trials required to meet the standards of parameter estimation. Two ADO algorithms, grid search and sequential Monte Carlo (SMC), are compared by examining the accuracy and precision of parameter estimation for the same model. The results showed that both algorithms performed similarly in terms of accuracy and precision. Grid resolution influences the performance of the grid search algorithm in such a way that insufficient grid resolution impairs efficiency while the gain diminishes as grid resolution increases. Given the satisfactory parameter estimation performance and faster computation, the grid search algorithm should be a favorable algorithm compared to SMC in situations of low dimensionality. This thesis also discusses several limitations of grid search in high-dimensional problems. Taking these results together, the present study has shown that GPU computing provides a feasible and effective solution for accelerating the computation of the ADO algorithm.