@mastersthesis{design-gsi-mastersthesis-20205,
author = "L{\'o}pez, Bruno Gonz{\'a}lez",
abstract = "The objective of this project is to train two intelligent agents to be able to solve text games,
in particular, a simple game inspired by ETSIT. In addition, metrics will be compared to
measure their performance in relation to a random decision making agent.
Text-based games (very popular in the 80s) are complex and interactive simulations
in which the text describes the state of the game and the players progress by introducing
textual commands. To solve this type of game, an intelligent agent must be able to explore
the environment, learn mechanics, identify its purpose, understand text and acquire some
temporal perception.
To achieve this, the project will be based on the machine learning field, specifically on
deep reinforcement learning that has proved to be the best solution for virtual environments
where agents have to learn to choose the best actions to achieve certain objectives. This
branch has become one of the most promising branches in the area of artificial intelligence.
The deep reinforcement learning algorithms DQN and A2C have demonstrated a great
performance in this kind of environments, so our agents are based on these architectures.
As games are a purely textual environment, the project will use tools from the field
of natural language processing. Understanding language requires skills such as long-term
memory, planning and common sense, qualities that our agent will need.
Textworld will be used for the management and creation of textual games. TextWorld is
a Python library developed by Microsoft that handles these games in addition to providing
tracking and reward control functions. It also allows us to create new games or automatically
generate others. The generative mechanisms it offers give precise control over the difficulty,
scope and language of the games built.",
address = "ETSI Telecomunicaci{\'o}n",
institution = "Universidad Polit{\'e}cnica de Madrid",
keywords = "TextWorld;Reinforcement Learning;Deep Learning;OpenAI",
month = "June",
title = "{D}esign and {D}evelopment of a {R}einforcement {L}earning-based {I}ntelligent {A}gent for {S}olving {T}ext {G}ames using {T}ext{W}orld",
type = "TFG",
year = "2020",
}