Research And Implementation Of Decision-making System In Adversarial Team Game

In the process of AI theory and technology applied, especially Multi-Agent theory in the military affairs and National Product Field, research of many problems must rely the emergence of multi-agent system. In recent years , a lot of adversarial team game system have become the focus of intelligent decision-making system, such as cooperative air combat system,multi-tank combat system and multi-robots soccer simulation platform,and so on. Because of their entertaining,confronting and dynamic, they were become the typical researchful platform for integration of Robot and AI technology. In these simulated system, they have the common characteristic, first, the tournament represent as a team combat with another team, each member is an self-determination robot. Secondly, during the tournament, each teammate not only need play self-technique and tactic capability, it is also need cooperating with other teammates. In order to winning the game, these simulated systems must integrate many technology and AI theory, including the structure of multi-agent system,the architecture of intelligent agent,decision-making system and game theory, machine learning, communication and so on. Between all these factors, decision-making system is the pivotal part of the whole system, this article focus to the decision-making system of adversarial team game system, this article coodinates from the team member's basic behavior tactics and team cooperation, designing the decision-making system with two layers, cooperative layer strategy and interaction layer strategy of sigle player. The decision-making system with two layers simplified the structure of decision-making system and strengthened inference capability, the main contents are as follows:Firstly, according to the role or the duty that cooperative layer strategy assigned, designing the basic behavior of interaction layer, this article utilizes java rule engine and genetic programming to design the hybrid action selector of interaction layer strategy. In this paper, we developed TableRex, which is a subsumption and behavior-oriented language, can be expressed on a fixed-length genome of genetic programming. With the advantage of the hybrid selector, improving the real-time and flexibility capability of interaction layer strategy.Secondly, with the foundation of interaction layer strategy, which provide the controller for each role or duty that assigned by cooperative layer strategy. By considering the development of Non-cooperative or cooperative reinforcement learning method based on the stochastic game in the Multi-Agent system and the game theory , we provided an algorithm: cooperative reinforcement learning method based on stochastic game in team-combat system, which learning the content of cooperative layer strategy. Meanwhile, the Q value of united-status and united-action of cooperative reinforcement learning method is expressed by an artificial neural.In this paper, we take Roboboce simulated combat system as the research object, designing the organization of robocode team and the architecture of robot, communication protocol and so on, then focusing the design and implementation of decision-making system,carrying on the examination separately to test the above contents.