Modelagem matemática e desenvolvimento de um agente baseado em modelos Markovianos

Abstract

In the last years, it is possible to see the advance of Artificial Intelligence applied to the most diverse games, whether they are deterministic or stochastic. Many techniques, despite achieving the expected result, require considerable computational power and are not viable on more basic platforms (e.g., low-cost smartphones). Thus, there is a need to create light weight techniques. This work uses Markov Chains in the development of an agent for a card game. The modeling was performed to represent the natural order of the forces and separates it into three modules according to the game’s modalities, being: Truco, Envido and Flor. Decision making is based on games played previously, reinforcing the decisions made and improving them. To evaluate the model, several rounds of tests were made, varying the opponent and the databases. This process resulted in learning matrices for each game mode. Thus, they were evaluated considering the main characteristics of the model and also of the opponents.