Моделирование Поведения Интеллектуальных Агентов На Основе Методов Машинного Обучения  в Моделях Конкуренции

Abstract

The article discusses the aspects of applying machine learning methods to existing methods for modeling the behavior of intelligent agents to enable agents to improve their performance in competition models.The practical significance of the study is represented by developing an approach to modeling the behavior of intelligent agents in order to increase the efficiency of their functioning in such areas as computer games, developing unmanned aerial vehicles and search robots, studying urban and transport mobility, as well as other complex systems.There is a review of the existing machine learning methods (reinforcement learning, deep learning, Q-learning) and methods for modeling the agents' behavior (a rule-based model, a finite automaton model of behavior, behavior trees). The authors have chosen the most appropriate combination of a learning method and a behavior model for the task: behavior trees and reinforcement learning.A test platform was implemented using Unity tools, behavior models were developed for the four main archetypes of agents that must compete in collecting resources in a limited time. A trained agent was implemented using Unity ML and TensorFlow tools.The test platform has become a basis for a series of experiments under various conditions: limited resources, resource abundance, average amount of resources. As part of the experiment, the authors tested the ability of the developed intelligent agent's behavior model to win in a competitive environment with agents equipped with various variants of traditional behavior models based on behavior trees. The efficiency and advantages of using the developed behavior model were evaluated. The paper analyzes the experimental results and draws conclusions regarding the potential of the selected combination of methods.