Learning Evaluation Function Based on Tree Strap in Shogi

IPSJ

Bibliographic Information

Other Title
  • 将棋におけるTree Strapに基づく評価関数の学習

Search this article

Description

近年,コンピュータ将棋では1万を超える評価項目を持った評価関数を用いることが一般的となっている.大量の評価項目それぞれに対し適切な重みを決定することは人手では非常に困難である.本論文ではコンピュータチェスにおいて成功を収めた自己対局を用いて評価関数を学習するTreeStrapを将棋に適用し,評価関数の学習を行った.実験では駒の重み,玉の安全度などを学習させ,得られる評価関数について考察を行った.また学習前の評価関数との対局実験を行うことでTreeStrapの有用性を調べた.実験の結果,学習前の評価関数に対し109勝26敗5分と勝ち越したことからコンピュータ将棋においても有用な手法であることが分かった.

Recently, many Shogi programs use evalation function that has many features. However it is difficult to determin many parameters of evaluation functions by heuristic. In this paper we apply Tree Strap method which learns evaluation functions using self-games and is succeeded in chess programs to Shogi. We used the weight of pieces, pieces in hand, the safety of the King and so on as features . We examin the learning the parameter of piece, piece in hand, the safety degree of the King and so on as parameters and discussed about the learned evaluation functions. We discuss the the learning result.In addition we performed the match experiments between the program which uses the learned parameters and the program which uses the old parameters in order to examine the effectiveness of the method. The program uses the learned parameters 109 wins, 26 loss and 5 draws against the program uses the old parameters. The results of the experiments show that the method is also useful for Shogi programs.

Journal

Details 詳細情報について

Report a problem

Back to top