Uma abordagem para melhorar o desempenho de agentes automáticos que operam em ambientes competitivos por meio de informações semânticas sobre mudanças de comportamento do oponente

Abstract

This Ph.D. work proposes and implements an unprecedented approach to increase the performance of agents operating in competitive scenarios that involve a data stream of information related to the dynamics of their adversaries’ behavior. Therefore, such agents must have the ability to detect, in real-time, eventual changes in the opponents’ behavior and, based on this information, adapt their decision-making processes in order to improve their abilities to deal with problems for which were designed. The RTS StarCraft: BroodWar game was used as a case study. The proposed approach was developed based on the following actions: 1) Extension and improvement of M-DBScan, which is a successful behavior change detection algorithm for data stream scenarios, in order to increase its accuracy of detecting behavior change, as well as being able to associate to each of these behaviors, a semantic that represents it; 2) Implementation of a StarCraft player agent whose decision-making module operates as follows: the information regarding the meanings of the opponent’s behavior provided by the extended versions of M-DBScan, proposed here, will be used to guide the agent in the execution of adequate actions considering the current opponent behavior. The proposed approach was validated through experiments conducted in order to evaluate the following metrics: accuracy in detecting changes in the opponent’s behavior and in assigning meanings to such behaviors; the win rate of the developed agent in tournaments where it faces different opponents in the game Starcraft. With the experiments carried out, it was possible to demonstrate the gain in the agent’s winning rate, due to the use of semantic information about the behavior change and the use of an adequate set of behaviors, which are relevant to the context of the problem.