Adaptiv Multi-Agent Förstärkningsinlärning för FPS-spel med hjälp av Djupa Q-nätverk

Sammanfattning

AbstractMost non-player characters (NPCs) in first-person shooter (FPS) games are still controlled by fixed, rule-based scripts. This makes them predictable, repetitive, and less engaging for players and reduces replay value. This project explores the application of reinforcement learning (RL) to train adaptable AI agents within a compact FPS. And it is specifically designed as SimpleFPS2v2. This project aims to teach RL agents, utilizing algorithms such as Deep Q-Networks (DQN). It is to exceed the fighting proficiency, adaptability, and strategic decision-making of conventional scripted bots. The environment replicates essential FPS mechanics, enabling systematic assessment of agent performance based on win rate, kill/death (K/D) ratio, and average team reward. Our results show that the RL agents can adapt to different opponent behaviors. It makes gameplay feel more dynamic and enjoyable. It is acknowledged further study is necessary to customize Artificial Intelligence (AI) behavior for specific players and more complex surroundings. But this work represents a significant experiment aimed at achieving more innovative, more responsive, and more enjoyable gaming experiences across many categories.