Comparison of Search Behaviors in Chess, Shogi, and the game of Go

Searching is important in games, and the relationship between depth and the best-changes has been investigated in chess programs using alpha-beta search. The combination of Deep Neural Network and Monte-Carlo Tree Search is a successful method in chess, shogi, and go, and it is important to investigate this method. Our purpose in this work is to find the differences in games and in search methods. If programs using the same method behaves differently in the different games, it may be possible to find the differences in the games. In this paper, we focus on an increase or decrease in evaluation value with increasing search depth reported in previous research in chess and the problem of fortress. We obtained the different results from the previous work when using Leela Chess Zero, a chess program with Deep Neural Network and Monte-Carlo Tree Search, although the results for other game programs with Deep Neural Network and Monte-Carlo Tree Search and other chess programs are the same from the previous work. The combination of a large number of draws and pUCT is a possible cause. We believe that there is room to allocate more resources to the best moves and that improvements can be made through ingenuity.