Immersive Game Worlds: Using Deep Reinforcement Learning for Lifelike Non-Player Characters

Research has shown that enhancing the believability of non-player characters (NPCs) in video games can increase player immersion. Traditionally, NPCs have been implemented with predefined behaviors through rigid coding. However, this approach leads to monotonous and repetitive actions as NPCs rigidly follow their programming, which affects their believability. Furthermore, the recent shift from 2D to 3D and first-person games has underscored the importance of believable NPCs. Therefore, it is essential to explore alternative methods for creating NPCs that can overcome the limitations of traditional approaches. Emerging artificial intelligence techniques, particularly machine learning and reinforcement learning, offer promising opportunities to replace traditional NPC development methods as NPC believability becomes increasingly important. In this study, we trained a shooter NPC using a deep reinforcement learning algorithm and compared it to another NPC that had the same goal but was developed using hard-coded instructions. The test results for both NPCs show that the model trained with reinforcement learning exhibits more realistic shooting behavior.