Monte Carlo Tree Search For "Dou Di Zhu"

The popular three-player game "Dou dizhu" in china is a typical incomplete information game.The game has not only confrontation between two farmers and the landlord,but also cooperation between two farmers.With the important progress of artificial intelligence technology in large-scale chess and card games,for example,Alpha Go and Libratus have achieved the ability to surpass current human professional players in Go and one-to-one non-limit Texas Hold'em games.The game "Dou dizhu" has also attracted widespread interest among artificial intelligence researchers.Monte Carlo Tree Search is an important method to solve the approximate solution of largescale search problems through the sampling method,it is an important part of Alpha Go Zero.In this paper,Monte Carlo Tree Search is combined with the characteristics of "Dou dizhu" and the convolutional neural network technology to explore and research the game strategy of "Dou dizhu".The specific work is as follows:1.Propose the Monte Carlo Tree Search Algorithm(MCTSHS)based on hand splitting.Statistics show that according to the rules of the game,when a hand is split into a card set that does not exceed the minimum number of splits plus 3,the card types chosen by human players in the actual game are included in this deck with a probability of over 99%.Therefore,on this basis,a hand-split algorithm is proposed to prune the game tree to improve its search efficiency.Experimental results show that the MCTSHS algorithm can make better game strategies without relying on human data.2.Propose the Monte Carlo Tree Search Algorithm(MCM)combined with convolutional neural network.Firstly,for MCTSHS,there is a problem that the decision-making time is too long and the learned strategy cannot be fully utilized,it is proposed to use the convolutional neural network to learn the historical decision data of MCTSHS,to realize the mapping of the state and possible decisions to the decision benefits,then choose the most profitable decision as the actual decision,and then greatly reduce the decision time;Secondly,in view of the errors in the learning of convolutional neural networks and the problem of MCTSHS's poor decision results due to experimental preset conditions,it is proposed to use Monte Carlo Tree Search Algorithm to improve the output of convolutional neural networks as the final decision.The experimental results show that the MCM algorithm is significantly better than the current mainstream "Dou dizhu" strategy.