همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

CAMQL-NPC: Enhancing Survival Behaviors in Dynamic Game Environments via Internal State Augmentation

نویسندگان :

Nima Salami¹ Hassan Haghighi²

1- Faculty of Computer Science and Engineering, Shahid Beheshti University 2- Faculty of Computer Science and Engineering, Shahid Beheshti University

کلمات کلیدی :

Non-Player Character (NPC)،Context-Aware Q-Learning،Internal State Augmentation،Survival Games،Dynamic Reward Shaping

چکیده :

Abstract—Designing intelligent Non-Player Characters (NPCs) in survival games requires agents to balance objective completion with self-preservation. Traditional Reinforcement Learning (RL) approaches, such as standard Q-Learning, often model the state space solely based on external environmental cues (e.g., distance to targets), neglecting the agent’s physiological status. This limitation leads to suboptimal behaviors, where agents pursue rewards despite critical health conditions, resulting in premature termination. This paper proposes a Context-Aware Modified Q-Learning (CAMQL) framework that integrates Internal State Augmentation (ISA). By incorporating the agent’s health status into the state representation and employing a dynamic survival-weighted reward function, the NPC learns to prioritize evasion over looting during critical scenarios. Experimental results in a Unity-based simulation demonstrate that CAMQL achieves a significantly higher convergence rate and survival probability compared to baseline MQL approaches, effectively bridging the gap between aggressive exploration and defensive preservation.