Online Planning on Low-Power CPU & GPU Systems on a Chip (SoCs)
This work proposes a new design for online planning for intelligent agents modeled as partially observable Markov decision processes (POMDP). This session introduces an online planner enhanced with Bloom filter memory that is implemented and evaluated on a low-power CPU and GPU system on a chip (SoC). Using the Data Parallel C++ (DPC++) parallel running model of the most compute-intensive kernel of this Bloom filter implementation, the overall planning time is reduced by 3.5 to 7.5 times for three representative benchmarks in the POMDP literature. Preliminary results promise new opportunities for using POMDP agents on low-power mobile platforms and in real-time use cases.
Denisa Constantinescu is a PhD candidate in the Department of Computer Architecture at the University of Malaga (UMA). She was a research visitor at the Northeastern University Computer Architecture Research (NUCAR) Laboratory in 2018. Her research interests are in low-power heterogeneous computing, intelligent control systems, and autonomous decision-making. She has been a mentor at Campus Tech Chicas UMA since 2018.
Rafael Asenjo is a professor of computer architecture at the University of Malaga. He has been using threading building blocks (TBB) since 2008, and over the past 10 years, he has focused on productively exploiting heterogeneous chips (CPU, GPU, and field-programmable gate array [FPGA]), using TBB as the orchestrating framework. He coauthored a book on TBB and is a member of the SYCL* advisory panel.