Research On Agent Cooperation In MAS Based On Game-Theory

Some issues which are related to the cooperation in Multi-Agent System are discussed in this paper. The concept of cooperation is re-interpreted in this paper firstly. It is defined as a kind of state of agent, not a kind of action of agent. An algorithm framework for modifying agent'belief based on HSN and an algorithm for tracking agent's intention based on FIRNN are described in this paper. Both of them are based on the basic model of agent, BDI model. Finally, it designs a simulation experiments for the research on the evolution of groups with multi-roles. The experiment is designed based on sociology society, and extends the traditional three Agent individual style.Game theory is a subject that studies the mathematical model which is used to express the conflict and cooperation between the rational decision-makers. The rational decision-makers are this kind of makers which will make decision according their individual preferences to make their benefit maximum. The most popular models are basic strategy model which is always used to express the static game and expansion model which is always used to express the dynamic game.MAS is broadly known as the agent-based calculation. Generally, it is defined as the system that is composed of the agents which are autonomous, intelligent, responsive, pre-action and social. MAS theory and technology of artificial intelligence originated from the traditional arising in the distributed artificial intelligence research. Currently, the research on theory and technology about MAS is gradually separated from the distributed artificial intelligence and integrated with some other fields. There are many outstanding effects when MAS is applied to many new areas, so it is the concentration in the academia and industry fields.Cooperation is one of the core issues in the research on MAS and interaction is the basis of the cooperation. The interaction between agents in the MAS is in essence a game process, so the current study focuses on making theory and method of game-theory as a tool to research the cooperation between multi-agents. In the game, the knowledge of the individual agent which comes from themselves and other agents will affect its decision-making and the state of the entire group.The knowledge held by agent is expressed as a belief in BDI model. In all intelligent systems that have the ability to learn, it is necessary to integrate the new knowledge and original knowledge. In MAS this process is known as Belief Revision. The initial method is the truth maintenance, but it only works in the isolated context. AGM, developed by Gradenfors, Alchourron and Makinson, solves this kind of problem. From then on, many new methods were proposed by the researchers, all of which are founded on the symbols/logic system and there are a lot of new problems coming. But what is worse is that these methods are only meaningful in theory except the sequence limit method and iterative method. It is very difficult to program for these methods. So it is necessary to explore new ways to express the agent's belief. In this paper, an algorithm framework for belief revision based on HSN is designed. In this algorithm framework, various type of information will be transformed into a type that can be accepted by neural network and saved in the neural network through the training. The approach could avoid the complex detection of conflict when integrate the new knowledge and original knowledge and be easy to program.The belief held by other agents is obtained by one agent through the course of the game either directly or indirectly. In the MAS expressed by BDI model, there is a problem that agent should know the whole knowledge if the agent gets belief with direct ways. So it is usually from the other side that to analyze the other's intention. Generally, it should build the other agent's BDI model firstly, then do reverse reasoning according the self-knowledge and other's action. The other idea is to predict the agent's intention according the action sequence that did by the special agent.Traditional multilayer feed-forward neural network is a static network, identified only applies to the handling, space curve fitting, and other issues. When faced such as time series prediction, dynamic system identification, etc., the need for a dynamic system itself for the neural network, the network is the need for short-term memory function. Usually there are two types: one is through delay element of the previous state of existence delay unit; another way is the introduction of a feedback network is a dynamic system. FIRNN (Finite Impulse Response Neural Network) is constructing the neural network in the first way. Structurally, FIRNN is composed of a group of sub-networks, which use the FIR filter layer as the connectivity between layers. At present, FIRNN has better effects in real-time speech recognition, multi-dimensional time series prediction, real-time control systems, as well as assess investment radar signal processing fields. Generally speaking, any fields which need to process variational data in space or time are all potential applications for FIRNN.The algorithm proposed in this paper is based on the second idea. This algorithm makes full use of the feature of FIRNN to recognize the window-size of data automatically. Then it builds many FIRNN with different delay time according the window-size have been recognised to predict the special agent's intention. This algorithm can effectively solve the problem of dividing time sequence. Finally, the algorithm is shown through a soccer match.It is the result of agent's cooperation for their benefit that the appearance of group. Traditionally, when doing research on evolution in groups, people set individuals as following types: Oxfam-good kind, altruistic and selfish-type. This paper extends these types and establishes an evolution simulation experiments with multi-role to study the evolution of groups。...