Aplicação de aprendizagem por reforço em um jogo educacional para educação ambiental

Abstract

Considering the growing relevance of educational games as pedagogical tools and the existing gap in the application of Artificial Intelligence techniques to personalize the learning experience, this work is justified by proposing an adaptive approach focused on environmental awareness. The objective is to present the development of the Recycle Learning game, which aims to teach the correct disposal of recyclable waste. This game stands out for its implementation of intelligent agents (non-playable characters) whose behavior is directly shaped by the player's actions. The game was developed on the Godot 4 platform and incorporates a mechanic in which the player's disposals, correct or incorrect, define the reward policy used in the learning process of the other agents. To justify the choice of the most suitable reinforcement learning algorithm for this dynamic scenario, a comparative analysis was conducted in a simulation environment developed in Python. In this environment, the Q-Learning (offpolicy) and SARSA (on-policy) algorithms were evaluated in scenarios with periodically changing disposal rules, simulating the variability resulting from the player's behavior. The results showed that Q-Learning performed better, achieving a higher success rate and greater stability (as indicated by less variation in the Q Table) compared to SARSA, especially in configurations with higher learning parameters. Therefore, it is concluded that the off-policy approach of Q-Learning is more robust and suitable for the proposed context, characterized by an unstable learning environment directly influenced by the player. Future work aims to apply the game in a school setting with students who have intellectual disabilities and in elementary school (grades 1-5). This undergraduate thesis summary was based on the article "Recycle Learning: An Educational Game with Intelligent Agents for Environmental Awareness," published in the proceedings of SBGames - XXIV Brazilian Symposium on Games and Digital Entertainment 2025.