Efficient Solving of Markov Decision Processes on GPUs Using Parallelized Sparse Matrices
Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Scientific › peer-review
Details
Original language | English |
---|---|
Title of host publication | 2018 Conference on Design and Architectures for Signal and Image Processing, DASIP 2018 |
Publisher | IEEE COMPUTER SOCIETY PRESS |
Pages | 13-18 |
Number of pages | 6 |
ISBN (Electronic) | 9781538682371 |
DOIs | |
Publication status | Published - Dec 2018 |
Publication type | A4 Article in a conference publication |
Event | Conference on Design and Architectures for Signal and Image Processing - Porto, Portugal Duration: 10 Oct 2018 → 12 Oct 2018 |
Publication series
Name | Conference on Design and Architectures for Signal and Image Processing, DASIP |
---|---|
ISSN (Print) | 2164-9766 |
Conference
Conference | Conference on Design and Architectures for Signal and Image Processing |
---|---|
Country | Portugal |
City | Porto |
Period | 10/10/18 → 12/10/18 |
Abstract
Markov Decision Processes (MDPs) provide important capabilities for facilitating the dynamic adaptation of hardware and software configurations to the environments in which they operate. However, the use of MDPs in embedded signal processing systems is limited because of the large computational demands for solving this class of system models. This paper presents Sparse Parallel Value Iteration (SPVI), a new algorithm for solving large MDPs on resource-constrained embedded systems that are equipped with mobile GPUs. SPVI leverages recent advances in parallel solving of MDPs and adds sparse linear algebra techniques to significantly outperform the state-of-the-art. The method and its application are described in detail, and demonstrated with case studies that are implemented on an NVIDIA Tegra K1 System On Chip (SoC). The experimental results show execution time improvements in the range of 65 % -78% for several applications. SPVI also lifts restrictions required by other MDP solver approaches, making it more widely compatible with large classes of optimization problems.
ASJC Scopus subject areas
Keywords
- CUDA, GPU, Markov decision processes, MDP, Sparsity, Value iteration