| 研究課題/領域番号 |
20K11932
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
福永 ALEX 東京大学, 大学院総合文化研究科, 教授 (90452002)
|
|---|
| 研究期間 (年度) |
2020-04-01 – 2023-03-31
|
| キーワード | 探索 / Heuristic Search |
|---|
| 研究実績の概要 |
Although the number of processing cores on CPUs and GPUs have continued to increase in recent years, recent work has shown that this increased parallelism does not necessarily result in improved performance of search algorithms. This project is an experimental and theoretical investigation of the tradeoffs between search efficiency and increased parallelism in search algorithms. We have analyzed greedy best first search algorithms for path-finding on graphs. We developed a theoretical framework to compare parallel best-first search with sequential best-first search, including both suboptimal (GBFS, Weighted A*) and optimal (A*) best-first search methods. We analyzed the extent to which the search behavior of existing parallel best-first search methods differ from sequential best-first search. We showed that existing methods are vulnerable to pathological behavior, and that they can expand nodes which would not be expanded by sequential search under any tie-breaking policy, resulting in arbitrarily worse performance compared to sequential search. We also proposed PUHF, a parallel best-first search which is guaranteed to expand a node only if there is some tie-breaking strategy for sequential search which expands the node.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
In the proposal, we planned to perform a theoretical analysis of parallel best first search in the first year. In the second year, we planned to develop new search algorithms which provide some theoretical bounds on the amount of search performed compared to sequential search.
We successfully analyzed parallel best-first search, showing that there exist search spaces where the search behavior of existing parallel best first search can be arbitrarily worse than sequential best first search. We also proposed a new variant of parallel greedy best first search which offers some theoretical
guarantees on the degradation relative to sequential search.
|
| 今後の研究の推進方策 |
We will continue the theoretical and experimental analysis and improvement of parallel search algorithms, focusing on bounded performance guarantees relative to seuential search.
|
| 次年度使用額が生じた理由 |
Spending of the reserach funds for this project was delayed due to the coronavirus pandemic, as much of the planned spending was for international travel, and international conference travel was not possible in 2021. The amounts carried over will be spent in the following year.
|
