Abstract: In the intelligent gaming field, EinStein würfelt nicht! gaming mobile policy should consider random dice rolls, chess piece positions, and potential interactions comprehensively, making the modeling and prediction more difficult. A strategy-building network based on the graph convolutional network (GCN) is proposed. A node selection method for the selection and simulation phases of Monte Carlo tree search (MCTS) is developed, and the situation features are extracted by multiple graph convolutional residual blocks and the probability of chess piece movement is output. By integrating the heuristic upper confidence bound for trees (heuristic UCT), the GP-UCT algorithm is formed based on the guidance of graph convolutional policy network. The experiment shows that the algorithm achieves at least a 60% win rate against the traditional UCT, BetaPrune and other baseline algorithms. The effectiveness of the proposed method in EinStein würfelt nicht! random gaming is verified.