Ensino de agentes inteligentes por meio de problemas em jogos

Abstract

The teaching of intelligent agents, presented in this study, was performed through the Problem Based Learning method. Pac-Man was selected as a case study. The constructive learning process was carried out through three stages, namely: (i) the implementation of basic agents, with the sole purpose of completing the route on the map; (ii) the implementation of real agents, considering the existence of ghosts; and, finally, (iii) the implementation of intelligent learning agents. As part of the knowledge building process, each new agent proposed by the student was created to solve problems found in the performance analysis of previous agents. In the first stage of development a performance improvement of 33.45% was observed from Agent 1 to Agent 6. In the second stage, considering the actual game, Agent 8 showed a performance increase of 20.49% when compared to Agent 7. In the third stage, artificial neural networks and genetic algorithms were used, which allowed us to create an agent capable of learning and completing the map alone. Thus, it was possible to prove that the selected techniques were efficient in improving the intelligence level of the agents proposed for the game in question. In addition, the use of this teaching method resulted in a greater involvement of the student with the Artificial Intelligence discipline, favoring the student's mastery in intelligent agent construction techniques, as well as contributing to his interest in this area of study.