Automatically Configuring Deep Q-Learning agents for the Berkeley Pacman project

Recently, there have been several advances on integrating Deep Neural Networks (DNNs) and Reinforcement Learning (RL) algorithms. These efforts led to the development of Deep Q-Learning (DQL) algorithms which have been applied successfully to develop competitive approaches for multiagent games. Both...

Descripción completa

Detalles Bibliográficos
Autor: Merino Pulido, Albert Eduard
Tipo de recurso: tesis de maestría
Fecha de publicación:2018
País:España
Institución:Varias* (Consorci de Biblioteques Universitáries de Catalunya, Centre de Serveis Científics i Acadèmics de Catalunya)
Repositorio:Recercat. Dipósit de la Recerca de Catalunya
OAI Identifier:oai:recercat.cat:10459.1/64813
Acceso en línea:http://hdl.handle.net/10459.1/64813
Access Level:acceso abierto
Palabra clave:Automatic configuration
Reinforcement learning
Deep Q-learning
Aprenentatge automàtic
Descripción
Sumario:Recently, there have been several advances on integrating Deep Neural Networks (DNNs) and Reinforcement Learning (RL) algorithms. These efforts led to the development of Deep Q-Learning (DQL) algorithms which have been applied successfully to develop competitive approaches for multiagent games. Both DNNs and RL algorithms are highly parameterized and di erent settings can have a dramatic impact on their e ciency. Thus, DQL algorithms can also greatly bene t from a good setting of their parameters. In this project, we show how to apply Automatic Con guration (AC) tools in order to explore efficiently the parameter search space. We have conducted an extensive experimental investigation in the Berkeley Pacman environment which con rms that AC tools can provide up to an additional 20% boost in performance to DQL agents.