Learning-based Multi-agent Negotiation Model

Agents should be flexible and autonomous during the process of negotiation, for the negotiation environment is open and dynamic. These features depend on the design of the agents' BDI models, which are finished before negotiating. If agents want to cope with the outer factors flexibly and make more profits, the update issues of these agents' BDI models must be done before negotiation begins so that the behavior choices of agents can be instructed by the negotiation models intelligently during the negotiation process. However, the current learning algorithms of the agent negotiation exist only in the process of the negotiation. The agents also lack for knowledge about their opponents before negotiation.A learning algorithm on the agent negotiation history is put forward in this thesis, which lay emphasis on the update issues of the agents' believes before negotiation. The action manner and decision-making model of the multi-agent system are described in this paper, where the learning schemes of the agent negotiation of Zeng and Sycara (1998) are referenced. Combined with the decision-making behavior contrast matrix given by Fatima and Wooldridge (2001), the learning mechanism and decision-making mechanism are further elaborated, which are embedded in the multi-issue negotiation model of multi-agent system. Feasibility analysis about enhancing the ability of the agent decision-making model is showed after the survey of the existing learning algorithms and decision-making models.Knowing more private information about their opponents can help agents increase efficiency and profit in the process of negotiation. With the analysis given above, the learning algorithm is demonstrated on the bilateral multi-issue negotiation history. The negotiation history from the service provider is classified first, followed by the definition of the negotiation thread and the negotiation model. Then the detailed implement scheme of the algorithm isshowed in the following sections.The algorithm aims at predicting the private information of agent's opponent. The predicted outcome is used to update the agent's belief before negotiating. Thus the agent can possess more information about the rivals. At the same time, the decision-making ability of the model is improved and the negotiation threads are shortened. Experiments are conducted to investigate the factors that affect the outcome of the algorithm. The experimental results show that the algorithm works well in conducting the agents' choices on behaviors and tactics during the course of negotiation.