AlphaGo's Decision Making

Abstract

In this paper, we study the similarities and differences between the process of decision making in humans and AlphaGo in playing Baduk (Go, Weiqi). Previous discussions of unique or unconventional moves of AlphaGo ignored how AlphaGo tends to play in different situations: (1) when AlphaGo is leading the game, (2) when she is falling behind, and (3) when the situation of the game is close enough. Nor did they pay due attention to the problem of strategic choice of moves of AlphaGo. We argue that (1) that AlphaGo tends to play very thick and safe enclosing moves when she is leading the game, (2) that she tends to play do-or-die (all-or-nothing or gambling) moves that are backed up by very carefully calculated scheming strategy, when there is no hope to win the game, and (3) that she tends to figure out creative moves in order to take the initiative, when the game is close enough. After sharpening the concept of strategy itself, we also argue that there is sufficient ground to ascribe strategic reasoning to AlphaGo. Based on DeepMind AlphaGo team's monumental paper in Nature [24] we will check to what extent our results are compatible with AlphaGo's structure and its operating principles. What is most striking in our examination of AlphaGo's decision making is that her features can be better explained by prospect theory [14] rather than by expected utility theory. In order to test this hypothesis, we analyze many examples from AlphaGo's games. We conclude by a brief discussion of the possible implications of the present study and the remaining urgent problems for future study.