| 研究課題/領域番号 |
17K00296
|
|---|
| 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) |
2017-04-01 – 2020-03-31
|
| キーワード | 探索アルゴリズム / 人工知能 |
|---|
| 研究実績の概要 |
We investigated GPU-based algorithms for satisficing graph search (algorithms which seek paths from a start node to a node which satisfies some goal conditions). In particular, we focused on GPU-based search algorithms for classical planning based on Monte Carlo Random Walk (MRW). We showed that straightforward parallelizations of MRW perform poorly.We developed Batch MRW (BMRW), a a generalization of MRW which performs random walk starting with many seed states, in contrast to traditional MRW which used a single seed state. We evaluated a sequential implementation of BMRW on a single CPU core, and showed that a sequential, satisficing planner based on BMRW performed comparably with previous state-of-the-art MRW-based planners. Then, we proposed BMRW/G, which uses a GPU to perform the random walks. We showed that BMRW/G achives significant speedups compared to BMRW and achieved performance competitive with the state-of-the-art satisficing planners on a number of standard benchmark domains.
|
| 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
As proposed in the plan of work stated in the previous year's status report, we focused this year on satisficing search (algorithms which seek feasible paths as quickly as possible, without guarantees of the optimality of the paths).
After showing that straightforward parallelization of previous Monte-Carlo Random Walk based satisficing planning performed poorly, we developed BMRW, a new parallel search strategy, implemented it on a GPU, and showed that it was competitive with state-of-the-art planning algorithms on a number of standard International Planning Contest benchmark problems.
|
| 今後の研究の推進方策 |
So far, we have focused on algorithms for the GPU. Next, we will attempt to integrate the GPU-based search algorithms we have developed with CPU-based parallel search algorithms into a unified, heterogeneous parallel search algorithm which effectively utilizes all processing resources. Continuing the work from this year, we will focus on parallel search for satisficing, classical planning.
|
| 次年度使用額が生じた理由 |
(理由）In this fiscal year, the majority of the planned and actual research expenditures we for travel to international conferences for the purpose of
presenting our research results from this project, as well as discuss ongoing work on this project with international researchers. Due to unpredictability of exact travel costs as well as the unexpected availability of another source of travel funds for the graduate student working on this project, we incurred a surplus.
（使用計画）We plan to use the surplus funds to pay for (1) travel expenses to present and discuss the research results with both international an national
researchers, and (2) additional CPU/GPU computing resources to perform large-scale experiments.
|
