I am very excited to share the current status of mahjong AI research with AlphaGo's Dongfeng.
one. What is the difference between Mahjong and Go?
From the perspective of game theory, Go is a complete information dynamic game, and Mahjong is a dynamic game with incomplete information. The information held by the two sides of the game in Go is symmetrical, while the information held by the opponents in mahjong is asymmetrical. Although everyone can see what cards each player played, you do not know what my hand is and I do not know what your hand is. The root cause of this information asymmetry is the randomness of the wall of cards.
The above differences between Go and mahjong (or chess and cards, Yi and Bo) determine the fundamental difference in their winning strategy. A chess game is essentially a Monte Carlo tree. The winning strategy is to choose or force the opponent to choose a branch. All the outcomes below this branch are self-victories. Only because of the many changes in chess, even AlphaGo is far from possible to traverse the entire tree, so AlphaGo will use the value network to estimate the probability of a winning outcome under a certain branch. In fact, human thinking is similar. It is always bad for opponents to kill a big dragon in Go, or to give opponents a car in Chess. Under these branches, the probability of winning the final game is greatly reduced.
The winning strategy for cards is to increase their own EV. I don't know what cards I want to touch next, or what cards my opponents have, but all possible situations make up a probability space. I only need to make a choice to maximize my scoring function's expectations. The difficulty of brand strategy is that there are too many factors that affect this probability space, and the specific effects are difficult to determine, such as the habit of playing cards.
The following figure is selected from the "Science Sparrow", the player's stupid card playing 10% risk of attack, the x-axis score for their own cards and cards, y-axis for the tour, z-axis for their own score expectations .
two. What is "winning"?
How to measure the player's competitive level? Is mahjong a "seven games three" game? In fact, regardless of whether you play chess or mahjong, it is not possible to rely on only one game to distinguish the player's competitive level. It is just like AlphaGo and Li Shih-chien are going to make a five-game matchup. When AlphaGo wins the first set, everyone does not feel that AlphaGo is necessarily stronger than Lee Seung-gyun. Go has three, five, and ten moves. For competitive mahjong (national standard mahjong, Japanese mahjong), distinguishing the competitive level of the top players requires at least 2,000 to 10,000 hands. (If you use a double-track system similar to a bridge, this number will significantly reduce).
Brushless DC Permanent Magnet Gear Motor
Brushless Dc Motor,Brushed Dc Motor,Industrial Machine Dc Motor,Brushless Dc Permanent Magnet Gear Motor
NingBo BeiLun HengFeng Electromotor Manufacture Co.,Ltd. , https://www.hengfengmotor.com