masterThesis
Beyond Star: um modelo de arquitetura de aprendizado para generalização de estratégias em jogos RTS
Fecha
2020-12-16Registro en:
ARAÚJO, Marco Antônio Silva e. Beyond Star: um modelo de arquitetura de aprendizado para generalização de estratégias em jogos RTS. 2020. 132f. Dissertação (Mestrado Profissional em Tecnologia da Informação) - Instituto Metrópole Digital, Universidade Federal do Rio Grande do Norte, Natal, 2020.
Autor
Araújo, Marco Antônio Silva e
Resumen
One of the main research fields under Artificial Intelligence, on the context of Digital
Games, is the study of Real-Time Strategy Games (RTS), which are commonly considered
the successors of classic strategy games such as Checkers, Chess, Backgammon and
Go, and impose great challenges to this area’s researchers due to the great complexity
involved. Currently, the field aims to study the RTS using StarCraft I and II as stage
of experimentation. The main feature sought in artificial agents developed to this kind
of game is high performance, having as its main objective to defeat specialist human
players. On this context it is inserted the generalization problematic, that is the capacity
of an artificial agent of reusing previous experiences, from different contexts, to a new
environment. Generalization is a very studied field by the scientific community, but still
poorly explored on the context of RTS. By this reason, this work proposes the Beyond
Star model, which consists in an architecture to generically represent the state-space of
Real-Time Strategy Games, using as base deep reinforcement learning techniques aiming
to learn effective strategies to be applied in several RTS environments. As a basis to the
architecture, it was developed a platform titled URNAI, a tool that integrates several
machine learning algorithms and several different game environments, such as StarCraft II
and DeepRTS. To analyse if the solution is capable of allowing agent generalization, tests
were carried out in DeepRTS and StarCraft II. The results demonstrate that the trained
agents were capable of generalizing their knowledge from one environment to the other,
showing a promising result that allows to validate this work’s proposal.