Aprendizaje de comportamientos en el juego de Pac-Man
Loading...
Official URL
Full text at PDC
Publication date
2025
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
El presente trabajo explora el uso de técnicas de aprendizaje supervisado para pre decir el comportamiento del agente Pac-Man frente a distintos escenarios del juego. Para ello, se ha utilizado como entorno el juego Ms. Pac-Man Vs Ghosts, sobre el cual se ha generado un conjunto de datos estructurado en función de intersecciones y movimientos. Se han entrenado y comparado modelos como MLP, TabNet y algoritmos clásicos de Scikit-learn, donde se han aplicado mejoras mediante ingeniería de características y análisis de importancia de variables.
El objetivo final ha sido no solo obtener un modelo preciso, sino también explicable, capaz de justificar sus decisiones en cada situación de juego.
This paper explores the use of supervised learning techniques to predict the behaviour of the Pac-Man agent in different scenarios of the game. For this purpose, the game Ms. Pac-Man Vs Ghosts has been used as an environment, on which a dataset structured according to intersections and movements has been generated. Models such as MLP, TabNet and classic Scikit-learn algorithms have been trained and compared, where improvements have been applied by means of feature engineering and variable importance analysis. The final objective has been not only to obtain an accurate model, but also explainable, capable of justifying its decisions in each game situation.
This paper explores the use of supervised learning techniques to predict the behaviour of the Pac-Man agent in different scenarios of the game. For this purpose, the game Ms. Pac-Man Vs Ghosts has been used as an environment, on which a dataset structured according to intersections and movements has been generated. Models such as MLP, TabNet and classic Scikit-learn algorithms have been trained and compared, where improvements have been applied by means of feature engineering and variable importance analysis. The final objective has been not only to obtain an accurate model, but also explainable, capable of justifying its decisions in each game situation.
Description
Trabajo de Fin de Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Curso 2024/2025
El código de este proyecto es accesible a través de este enlace:
https://github.com/beja28/Ms.Pacman-Machine-Learning