Design And Implementation Of Card Game AI System Based On Supervised Learning

Artificial Intelligence(AI)technology is rapidly evolving and showing great potential in the gaming field.Game AI can be divided into two categories: one is aimed at defeating the top human masters,which will consume a lot of arithmetic power;the other is aimed at becoming an ordinary human playmate,which focuses on ensuring a certain level of play while reducing arithmetic power,like the DouDiZhu AI.In order to maintain the winning rate of the AI while effectively reducing its arithmetic overhead,the relationship between the winning rate and the number of remaining hands was investigated,and a new modeling approach was proposed based on the law that the stronger the effect of the handing strategy on the winning rate when the number of remaining hands is small.Specifically,the game of DouDiZhu is divided into two stages according to the game progression,the normal game stage and the residual stage.The division is based on the following criteria: When a player has no more than two cards in his hand,the game moves from the normal game phase to the handicap phase.During the normal game phase,the system only uses the model trained by the Convolutional Neural Network(CNN)to play cards,taking full advantage of the CNN's advantages in spatial feature extraction and effectively reducing the arithmetic power consumption.In the handicap stage,the system first uses CNN-trained card guessing model to calculate the player's remaining hand,then pushes back another player's hand based on the card information,and finally constructs the game tree with the three-way hand information and search for the best solution,effectively increasing the game's winning rate while controlling the power consumption.In this article,the AI system of this DouDiZhu game is trained and tested,and the system is evaluated in terms of the accuracy of the playing model,the accuracy of the guessing model,the winning rate,the overhead,etc.Under the experimental conditions,the overall win rate of AI against people was 67% for landowners and 51% for farmers,with an average calculation time of 17.31 ms per step,proving that the system has a certain overall win rate,and at the same time has excellent performance,reaching an acceptable level for commercial applications.