Seleção de atributos para modelagem preditiva de estratégias em jogos de Pôquer

Abstract

Poker is a card game with a very simple objective: players bet that their card combination is better than the ones from their opponents. Given its characteristics, like competitive multi-agent environment, partially observable and stochastic, the game is used as a research problem in the areas of Artificial Intelligence and Machine Learning. The problem of learning a player’s strategies through of his previous gamelogs was recently addressed in the literature. In that work, a dataset was modeled to represent four aspects of the game: position, aggression, hand quality and current situation. Some classification techniques had a satisfactory result, but the importance of the attributes was not analyzed. Since redundant or irrelevant attributes may affect the predictive performance of the classifiers, this work aims to analyze how does the attribute selection process may help to deal with the problem. The ideal approach to select the best subset of attributes would be to try out all possible subsets, but it is an unfeasible method due to the large amount of possible subsets. Thus, we investigated in the literature methods for selecting attributes such as filters and wrappers to be used in a computational environment reproduced from the literature and adapted to perform attribute selection. It was identified that certain attributes are important in certain stages, but can become irrelevant during the game. There were filters that stood out both predictive performance and attribute reduction, but the amount of attributes removed is not always related to the improvement of performance. The wrapper approach proved to be more effective, in both points (attribute reduction and predictive performance), although requiring a higher computational cost. In general, the selection of attributes demonstrated to be an important step to data pre-processing, as there was improvement in several classifier in all stages of the game.