圍棋布局、中盤、官子階段之認知能力：當實驗心理學遇上人工智慧

Cognitive Abilities in the Game of Go during the Opening, Middle, and Endgame Phases: When Experimental Psychology Meets Artificial Intelligence

圍棋被認為是最複雜、最具挑戰性之棋類運動，其進行過程大致分為三個階段，分別為前期的「布局」、中期的「中盤」及後期的「官子」。雙方棋士皆需運用大量的認知處理能力，包含決策判斷、推理計算等，方能一步一步爭取棋局之最終勝利。然而過去心理學研究對於圍棋認知處理著墨甚少，未有理論探討圍棋三階段所需運用之認知能力是否相同或相異。本研究以實驗心理學研究方法搭配人工智慧之分析技術，探討圍棋三階段之認知能力。本研究由職業棋士編製不同階段之圍棋題目，使題目難度約為業餘棋士高段位之棋力，並且讓棋局符合實驗心理學之基本要求。研究參與者為24名業餘棋士，在電腦上進行不同階段的圍棋題目，並搭配空間視覺搜尋、邏輯推理及數學計算三種認知干擾作業，以探討這些干擾作業對於各階段圍棋題目之認知處理的影響。實驗結果發現，空間干擾作業會顯著影響布局階段之正確率，但對於中盤、官子階段並無影響；推理干擾則會顯著影響中盤、官子階段之正確率，對於布局並無影響；而計算干擾對於三階段圍棋題目的影響並未達到統計顯著。另一方面，本研究亦運用人工智慧之技術，以開放原始碼之深度學習圍棋程式將所有實驗題目進行分析，大量模擬並計算每個選項的勝率，發現與題目正解相當一致。並且利用人工智慧中機器學習演算法（隨機森林演算法、支持向量機演算法與深度神經網路）搭配置換檢驗法，對於棋士們的正確率與反應時間進行分析。結果顯示這個演算法能夠將實驗數據有效地區辨為三階段之圍棋題目、或是區辨為不同認知干擾類型，其預測正確率遠高於隨機猜測水準及虛無假設分配之決斷值，且經由交叉驗證程序可以證實此結果具有良好之可推論性。總結而言，本研究顯示圍棋不同階段所需之認知能力不完全相同，而人工智慧分析也進一步支持這個論點。在對弈過程中的布局階段需具備大局思維，因此空間能力是影響布局階段時棋局判斷之重要因子；而較需沙盤推演之中盤與官子階段則主要依賴推理能力。本研究結果不但有助於瞭解圍棋對弈過程所運用之認知能力，也藉此將人工智慧與實驗心理學融合運用於探討人類智慧與心智功能，期盼未來人工智慧能對於心理學研究帶來更大的影響與貢獻。

The game of Go, also called Weiqi or Baduk, is one of the most sophisticated board games in the world. The players compete with each other by surrounding more territory using their stones. There are three phases in a game of Go, including the opening, middle, and endgame. Only a very few psychological studies have investigated the underlying processes or neural mechanisms used while playing Go. It has been suggested that some cognitive abilities may be important during the game, but the recruitment of different kinds of cognitive abilities in three phases is still unknown. The present study addressed this issue by combining experimental psychology approaches and artificial intelligence (AI) algorithms. Twenty-four Go players tried their best to quickly answer 48 Go questions in each of three phases, with different cognitive interference tasks appearing simultaneously. Their accuracy and reaction time on these questions were recorded as their performance. The Go questions were designed and organized by a professional Go player, and some basic requirements for psychological experiments were met. In addition to a control task, there were three types of interference tasks: a visual spatial search, logical reasoning, and calculation. The results showed that the spatial interference task decreased the accuracy in the opening phase, suggesting that spatial ability is the most important cognitive ability used in the opening of a Go game. The logical reasoning interference task decreased the accuracy in the middle and endgame phases, implying that reasoning ability is very critical in these phases. The calculation task had a less significant interference effect. In addition, we used three AI-related algorithms to classify the subjects' performance in the three phases of Go questions under different degrees of cognitive interference. The results showed that these algorithms had much better than chance accuracy to correctly classify the performance in three different phases of Go questions or under different degrees of cognitive interference. Cross validation procedures ensured the generalizability, and permutation tests also indicated that the predictive accuracy of these models was statistically significant. We thus argue that there are indeed different cognitive representations in these three phases under different levels of interference. In summary, in the present study, an experimental approach was adopted to reveal the involvement of cognitive abilities in three phases of Go. In addition, we provide a new perspective for experimental psychology by introducing an AI-related analysis of multivariate data, which infers that artificial intelligence can have a greater influence and make a greater contribution to the understanding of psychology and human intelligence.