Publication: Aprendizaje por refuerzo: Fundamentos teóricos del algoritmo AlphaZero e implementación
Loading...
Official URL
Full text at PDC
Publication Date
2021
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
En 2016, el equipo de DeepMind sorprendió al mundo creando una inteligencia artificial capaz de jugar al go a un nivel superior al de los humanos y ganando a uno de los jugadores más laureados de la historia. Sin embargo, AlphaGo era un algoritmo complejo y requería de una gran potencia computacional.
Un año más tarde se publicó AlphaZero. La belleza de este algoritmo residía no solo en que requería menos potencia computacional y se podía aplicar a más juegos, sino en la elegancia con la que combinaba sus componentes para lograr un rendimiento por encima de cualquier otro algoritmo hasta el momento.
El objetivo de este trabajo es explicar el funcionamiento del algoritmo AlphaZero. Para ello se introducen primero las nociones teóricas básicas del aprendizaje por refuerzo y las redes neuronales y posteriormente los detalles particulares del algoritmo. Además, se implementa una versión reducida del mismo y se entrena para jugar al tres en raya y al Conecta 4, estudiándose los resultados obtenidos.
In 2016, DeepMind’s team surprised the world by crafting an artificial intelligence that was able to play Go at a superhuman level and win the second most laureate Go player in history. However, AlphaGo was a complex algorithm, that required huge computing power. A year later AlphaZero was published. The beauty behind this algorithm relies not only on the smaller computing power required or that it can be applied to more board games but also on the way they skillfully put together its components to achieve a performance way better than other Go programs at that moment. The objective of this work is to explain how AlphaZero works. First, we briefly introduce the theoretical basis of reinforcement learning and neural networks and later we explain the details of the algorithm. In addition, a slightly simplified version of the algorithm is implemented and trained to play Tic Tac Toe and Connect 4, and its performance is analyzed.
In 2016, DeepMind’s team surprised the world by crafting an artificial intelligence that was able to play Go at a superhuman level and win the second most laureate Go player in history. However, AlphaGo was a complex algorithm, that required huge computing power. A year later AlphaZero was published. The beauty behind this algorithm relies not only on the smaller computing power required or that it can be applied to more board games but also on the way they skillfully put together its components to achieve a performance way better than other Go programs at that moment. The objective of this work is to explain how AlphaZero works. First, we briefly introduce the theoretical basis of reinforcement learning and neural networks and later we explain the details of the algorithm. In addition, a slightly simplified version of the algorithm is implemented and trained to play Tic Tac Toe and Connect 4, and its performance is analyzed.
Description
Trabajo de Fin de Grado en Doble Grado en Ingeniería Informática - Matemáticas, Facultad de Informática UCM, Departamento de Sistemas Informáticos y Computación, Curso 2022-21.