Casos e clusters no desenvolvimento de políticas de reuso para tomada de decisão em jogos de cartas

Abstract

This work investigates combinations of cases and clusters that use reusable gaming actions present in retrieved cases from queries made by virtual players. These queries are built using case-based reasoning (CBR) designed for playing cards action and bet actions. With the support of the K-MEANS data clustering algorithm, the past experiences are organized in groups where the combination of game-state and game-actions is considered. Initially, a criterion is applied to choose the group to be accessed; then, another reuse criterion uses the cases that belong to the selected cluster to determine which game-action should be reused as a solution to the current game situation. This two-step reuse model was implemented with different combinations of criteria in different steps. The criteria explored in this work are identified by majority rules reuse, choice of action/cluster by random methods where the probabilities of each action being classified are considered, actions/clusters most likely to win, or choice of cluster or group that provides the most points. To evaluate these combinations, the virtual agents were implemented with different reuse policies and were used in different tests. To validate the agents, this work shows approaches to all stages of the CBR cycle (retrieval, reuse, revise, retain).