Combining NEAT and PSO for learning tactical human behavior

摘要

This article presents and discusses a machine learning algorithm called PIGEON used to build agents capable of displaying tactical behavior in various domains. Such tactical behavior can be relevant in military simulations and video games, as well as in everyday tasks in the physical world, such as driving an automobile. Furthermore, PIGEON displays good performance across two different approaches to learning (observational and experiential) and across multiple domains. PIGEON is a hybrid algorithm, combining NEAT and PSO in two different manners. The investigation described in this paper compares the performance of the two versions of PIGEON to each other as well as to NEAT and to PSO individually. These four machine learning algorithms are applied in two different approaches to learning-through observation of human performance and through experience, as well as in three distinct domain testbeds. The criteria used to compare them were high proficiency in task completion and rapid learning. Results indicate that overall, PIGEON worked best when NEAT and PSO are applied in an alternating manner. This combination was called PIGEON-Alternate, or simply Alternate. The two versions of the PIGEON algorithm, the tests conducted, the results obtained and the conclusions are described in detail.