• 検索結果がありません。

MPI による並列モンテカルロ木探索の実装と評価 要旨

N/A
N/A
Protected

Academic year: 2021

シェア "MPI による並列モンテカルロ木探索の実装と評価 要旨"

Copied!
2
0
0

読み込み中.... (全文を見る)

全文

(1)

要 旨

MPI

による並列モンテカルロ木探索の実装と評価

藤本 佑馬

モンテカルロ木探索は,多人数ゲームや二人ゲームだけでなく一人ゲームなど様々なゲー ムに有用なアルゴリズムである.またプランニングやバイオメトリックスシステムを用いた セキュリティ評価などにおいても有効であることが示されており,汎用性が高い.モンテカル ロ木探索において, シミュレーションの回数を増加させることで精度の高い探索を行うこと ができる.しかし,シミュレーションの回数を増やすことは全シミュレーション終了までの実 行時間の増加につながる. この問題を解決するため本研究では, 並列プログラミングの規格 のひとつである Message-Passing Interface を用いて,モンテカルロ木探索を並列化する. れによってシミュレーション回数を増加させながら,全シミュレーション終了までの実行時 間の短縮を図る.並列モンテカルロ木探索を実現するために Root 並列化と Tree 並列化の 2種類の手法を使用した. 16台構成のクラスタを用いて実験を行った.結果,プロセス間の通 信回数を調整することで高速化することができた.

キーワード Message Passing Interface, モンテカルロ木探索, 並列化

i

(2)

Abstract

Implementation and Evaluation of

Parallel Monte-Carlo Tree Search with MPI

Yuuma Fujimoto

The Monte-Carlo tree search is an algorith useful in not only two-person games bust also one-person games and multiplayer games. Moreover, it is a highly flexible algorithm as its effectiveness is shown also in plannings, the security evaluation using a biometrics system, and so on. In the Monte-Carlo tree search, we can search a more precise answer by increasing the number of simulations. However, increasing the number of simulations leads to the increase of the execution time. In order to solve this problem, in this study, the Monte-Carlo tree search is parallelized with the Message-Passing Interface that is a widely-used standard of parallel programming. By parallelizing the algorithm, we shorten the execution time while making the number of simulations increase. In order to realize parallel Monte Carlo tree search, two techniques, Root parallelization and Tree parallelization, were used. Experiments were conducted on a cluster of 16 computing nodes. With a suitable adjustment of communication frequency, the overall computation was accelerated.

key words Message Passing Interface, Monte-Carlo Tree Search, Parallelization

ii

参照

関連したドキュメント

仮想並列計算の構築には、 MPI(Message Passing Interface) [1][2] や PVM(Parallel Virtual Machine) [5][6] 等のソフトウェアを用いて行うことができる。MPI

A new UCT search method using position evaluation function and its evaluation by Othello Shota Maehara,†1 Tsuyoshi Hashimoto†2 and Yasuyuki Kobayashi†1 The Monte Carlo

Abstract: This paper proposes a modified Monte-Carlo tree search (abbreviated as MCTS) for playing 19 ×19 go. Diversifying playout causes the MCTS to have behavior just

In this study, we have modified the source code of dlshogi to create a program that combines Monte Carlo softmax search and evaluation functions of neural network

including the game of Go. In this paper, we present a learning method for the playout policy in Monte-Carlo tree search based on the Simulation Balancing method. In our

Asynchronous parallel game-tree search methods are effective ways to improve the playing strength by utilizing many computing nodes connected in low-cost network systems..

Abstract: Monte Carlo Tree Search MCTS is a search method for perfect information games and can be effectively used for incomplete information games too.. Several extensions for the

CUDA implemantation of a currency option valuation using a parallel programming framework.. Kazutaka Toyabe†1