A Player-like Agent Reinforcement Learning Method For Automatic Evaluation of Game Map

摘要: Game map is an important human-computer interactive content-bearing platform in major games. With the application of cellular automata(CA) and Procedural Content Generation (PCG)in map generation, the spatial scale and data volume of current game maps are increasing greatly, while in game map test procedure, automatic methods such as interactive test script are inadequate both in depth and application breadth, especially in the lack of game map evaluation from player experience perspective. This research proposes an automatic game map test method based on agent reinforcement learning. By establishing agents’ interactive action models standing for different types of players’ behaviors in the map, universal evaluation of the map environment is enhanced through agent actions, which can optimize game map design from the perspective of player experience with quantitative value of inferiority. Finally, our campus scenes in Minecraft were used as the experimental environments to verify the effectiveness of the method.