Pengembangan Automated Prompt Engineering untuk Bot-Companion NPC dalam Game Edurobots

Abstract

This research focuses on the development of an Automated Prompt Engineering (APE) system for a bot-companion NPC in the EduRobots game, aiming to design a system capable of automatically generating prompts based on in-game conditions so that the bot-companion can respond adaptively and contextually. The system utilizes a 3x3 grid sensor that detects key objects within the game area, such as the player, enemies, coins, medkits, and artifacts. The detected data are then processed by the APE system to construct dynamic prompts representing the game situation contextually. These prompts are sent to the Large Language Model (LLM) GPT 4o mini through an API to obtain decisions or actions that the bot-companion will execute. Several gameplay scenarios, both with and without additional instructions from the player, were tested to evaluate the system’s ability to interpret the game state and generate appropriate responses. The results show that the developed grid sensor and APE system perform effectively in detecting game conditions and generating accurate prompts. The GPT model successfully produced relevant responses based on the game context and adapted its actions according to the player’s instructions. Furthermore, the system demonstrated adaptive behavior when facing situations that did not match the given instructions. Therefore, the implementation of APE in this study proves to enhance the intelligence and adaptability of bot-companions.