電腦圍棋的連接性之研究

Abstract

本論文的研究方向是尋找及辨識圍棋盤面中棋子之間的連接關係，藉以提供圍棋程式更多資訊並幫助解決死活問題。為了判斷棋子間的連接性，我們從小範圍區域中的兩顆棋子間關係開始分析，並遞回擴展到整個盤面的棋子間關係，並利用這些關係，在MCTS中加入兩個connection相關的政策，以分析是否加強MCTS的選步能力。 實驗顯示，加上這些政策後，在每步棋同樣模擬數的情況下，與GnuGo 3.8 level 10對戰，提升約3%到6%的勝率，但效能大幅降低，這部分未來需要更進一步改善。

In this thesis, we propose a method which can identify and analyze various stone relations in Go positions. This information can be used by Go program. To achieve this we start by analyzing the relations between two stones, then recursively expand the relations to the stones over the entire position. We then add two “connection policies” to MCTS in order to see if they increase the capability of choosing moves. Experiments show that by adding these policies, we achieved an increase of 3% to 6% in win rates against GnuGo 3.8, level 10, with the same number of simulations per move. However, the computation time needed for a single simulation largely increased. The performance improvement is critical in the future.