A model for automatic agent generation based on machine learning in simulation games

Abstract

With the continuous promotion of Artificial Intelligence (AI) technology in games, researchers are generally willing to pay attention to the effect of AI on game picture quality and man-machine games. However, how to generate a large number of non-player characters (NPC) in a game, especially in a simulation game, is still a problem. Even if some people use machine learning to discover solutions, most answers tend to be limited to copying finite-state machines established by programmers. This paper explores techniques, such as machine learning and maze algorithms, for 2D in-game environments to construct a model that can quickly generate NPCs with 'personality' and adapt to the environment when developing a game. In the process of model construction, the temporal difference algorithm and the depth-first algorithm are used to customize the decision structure for the NPC. At the same time, a simple node-style map generation algorithm is added to restrain NPC. Moreover, an optional neural network model gives the NPC a memory that can be updated. This model successfully realizes the rapid construction of NPC in developing some simple 2D games. In addition, NPCs have a fair amount of intelligence. It can be used during game development or while the game is running and generate a certain number of NPCs in an unfamiliar environment with differences in behavior.