Abstract
目前大多數的圍棋程式仿效人類下棋的思路,嘗試將專家的知識轉為電腦系統.然而將專家知識轉換為電腦系統需要大量人工輔助,同時存在於心中的知識也不易精確且完整的轉化為電腦系統. 在本論文中,我們提出數種用演化學習方式去增強圍棋程式的方法,同時我們設計了數個電腦棋手.這些電腦棋手能夠以演化學習法自動增進棋力,不需要大量人工輔助,而學習後的電腦棋手的內部知識也提供我們對圍棋知識的進一步了解.我們設計了一個稱之為位置樣式演化棋手以位置樣式規則表示其下圍棋的策略.應用遺傳演算法,位置樣式演化棋手能夠逐步增加其棋力.而將位置樣式的知識加入固定策略的棋手知識中,並用遺傳演算法學習,可以擊敗原本的固定策略棋手. 另外,我們設計了一個稱之為簡單訣竅的固定策略棋手.簡單訣竅棋手應用兩個簡單訣竅下圍棋.這些訣竅的參數是由我們所設定的.為了表現出合理的策略,這些參數需要經過許多次的試誤與調整,而且我們所設定的參數可能並非最佳的參數.因此我們用演化學習法去調整這些訣竅的參數.實驗的結果顯現出,經演化學習後棋手能表現的比原本我們所所設定的棋手要好.我們也嘗試分別學習簡單訣竅棋手的兩個訣竅,再將分別學習到的訣竅組合起來.實驗的結果也顯示,將訣竅分別學習也能表現出不錯的結果,能勝過我們原先所設定的棋手.Most computer Go programs try to simulate the thought ofhuman players in Go game and code the expert knowledge to thesystems. It needs a large amount of human efforts and is noteasy to transfer expertise into a system in a complete andaccurate way. In this thesis, we develop several methods usingthe Genetic Algorithm (GA) to increase the strength of Goprogram. We design several kinds of artificial players andconduct several experiments to show that we could use thesemethodologies to increase the abilities of Go program. Theposition pattern evolving players (PPEP) use position patternrules to represent the policies of playing Go games. Using GA, aPPEP could increase its strength. We design a fixed-policyplayer called simple heuristics player (SHP) A player combininga PPEP with a SHP will outperform the original SHP. The SHP usetwo simple heuristics to play Go game. To tune the parameters ofthe heuristics manually may take a lot of efforts and the resultmay not be optimal. We use evolutionary learning to tune theseparameters automatically. The results show that the player usingevolutionary learning could outperform the original SHP. Wecould also learn the two heuristics separately and then combinedthem. The player trained this way will still outperform theoriginal SHP.