Análise de viabilidade de utilização de deep q-learning para criação de um controlador no torcs

Abstract

Aiming at the growth of the autonomous car and game industries, the feasibility analysis of an artificial intelligence algorithm plays a key role in the development of new games and cars. This paper analyzes the feasibility of using Deep Q-Learning using TORCS, an open source car racing simulator. Changing a reward function was tested during development and showed promising behavior for the start of training. For this, efficiency and learning tests are proposed. It was not possible to reach a complete controller using only Deep Q-Learning, so it was not feasible to use. A discussion is made about the results achieved and the proposed changes are compatible with models already used in other works.