2019 Fiscal Year Research-status Report
ヘテロジニアス計算機環境における並列探索アルゴリズムの研究
| Project/Area Number |
17K00296
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
福永 ALEX 東京大学, 大学院総合文化研究科, 教授 (90452002)
|
|---|
| Project Period (FY) |
2017-04-01 – 2021-03-31
|
| Keywords | 探索アルゴリズム / 人工知能 / 並列アルゴリズム |
|---|
| Outline of Annual Research Achievements |
We initially investigated heterogeneous graph search algorithms which used both the CPU and GPU, as initially planned. However, we discovered that for satisficing search, in which we seek (non-optimal) solutions to difficult path-finding problems, we discovered that before considering heterogoneous CPU-GPU algorithms, there were fundamental challenges in parallelization on CPU-only systems which needed to be addressed.
Although A* search can be efficiently parallelized using methods such as Hash-Distributed A* (HDA*), distributed parallelization of Greedy Best First Search (GBFS), a suboptimal search which often finds solutions much faster than A*, has received little attention. We found that surprisingly, HDGBFS, an adaptation of HDA* to GBFS, often performs significantly worse than sequential GBFS. We analyzed this performance degradation, and proposed a novel method for distributed parallelization of GBFS, which significantly outperforms HDGBFS on a set of standard domain-independent planning benchmark problems.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The original plan was to develop heterogeneous parallel graph search algorithms using both the CPU and GPU. In the first 2 years, we focused primarily on GPU-based search algorithms, and successfully developed new GPU algorithms for both optimal search (described in our 2017 Symposium on Combinatorial Search paper) and a satisficing search (described in our 2018 International Conference on Automated Planning and Scheduling Paper). The surprising discovery that satisficing, parallel best-first search often performs dramatically worse than sequential search on some benchmark problems caused us to focus on CPU-based search algorithms this year, so although we were unable to develop a new heterogeneous algorithm as originally planned, our work this year resulted in significant new insights on CPU-based parallel, satisficing search (described in our 2019 ICAPS paper).
|
| Strategy for Future Research Activity |
This was the final year of the project. In a future project, we will continue the investigation of parallel best-first search algorithms, focusing on a better theoretical understanding of the properties of parallel best-first search.
|
| Causes of Carryover |
We planned to use some funds to attend an international conference in early 2020, which we were unable to attend, resulting in unused funds. We will use these carryover funds to attend another international conference (e.g., International Conference on Automated Planning and Scheduling).
|
