The project aims to investigate the integration of Large Language Models (LLMs) into gaming environments, focusing specifically on the development of intelligent agents. The advent of sophisticated LLMs such as GPT-4 offers unprecedented opportunities for enhancing game interactivity and realism by enabling agents to understand and navigate complex environments, solve puzzles, and interact in a human-like manner. Despite the potential, the application of LLMs in real-time gaming presents significant challenges, including computational demands, real-time decision-making, and the complexity of game environments. This research aims to bridge this gap, exploring the feasibility and methods for deploying LLM-based agents within interactive gaming environments focusing on maze-solving tasks.

The study adopts a comprehensive approach, evaluating various LLM architectures and configurations across two distinct environments: a simplified maze and the more complex Obstacle Tower environment. Through extensive testing, the research assesses the agents' navigational abilities, problem-solving skills, and interaction capabilities, alongside examining LLM response latency, models' pre-existing knowledge of the game environments, and spatial reasoning skills.

The findings reveal that LLM-based agents, particularly those utilizing GPT-4, exhibit promising capabilities in maze navigation and puzzle-solving, outperforming other models (e.g., GPT-3.5, Mixtral-8x7B, Llama3-13B) in both test environments. However, the study also highlights the limitations related to LLM latency and the challenges of integrating these models into real-time game applications.

This research opens the potential for insights into broader applications of LLMs in interactive and dynamic settings. By demonstrating the potential of LLM-based agents in gaming, this thesis lays the groundwork for future innovations in creating more immersive and intelligent game experiences, as well as advancing the application of natural language processing in novel contexts. Furthermore, it opens avenues for further research into optimizing LLM performance in real-time environments and exploring new applications for LLM-based interaction in interactive systems.

Detta projekt syftar till att utveckla effektiva och kostnadseffektiva AI-agenter för spelapplikationer, vilket kan bidra till att minska utvecklingskostnaderna och förbättra användarupplevelserna. Genom att utnyttja förtränade modeller kan vi skapa en mer engagerande spelupplevelse. Dessutom undersöker vi potentialen hos LLM-agenter i spelomgivningar, med fokus på deras förmåga att fungera som högnivåplanerare. Detta arbete bidrar till att belysa de relevanta ekonomiska, sociala, miljömässiga och etiska aspekterna av AI i spel.