Research On The Application Of Reinforcement Learning And Opponent Classification In Human-Agent Negotiation

The rapid development of the Internet has led to the prosperity of the e-commerce industry.With the increasing frequency of online business transactions and the increasing scale of transactions,online negotiations and communication between sellers and consumers have become more and more frequent and close,so realizing the automation of online transaction negotiation process has gradually become the potential demand of enterprises or individuals.The development of artificial intelligence technology provides an opportunity for the realization of the automated negotiation system.Using the Agent technology can delegate the users to deal with complex negotiating matters,it could enable users negotiate with multiple parties anytime and anywhere,as well as greatly improve the negotiation efficiency and reduce the cost of sales.Therefore,Agent-based automated negotiation system has become an important research direction for many scholars.At present,the research on the automated negotiation system mainly focuses on the negotiations between Agent and Agent.There are relatively few researches on Human-Agent negotiation,and the researchers often neglect classifying negotiating opponents,hence cannot adopt appropriate negotiation strategies according to the opponent's type,resulting in low negotiation performance.Therefore,based on the theory of reinforcement learning negotiation,this thesis introduces the adaptive learning mechanism of the opponent type to improve the negotiation effect,and builds a Human-Agent negotiation model.This thesis has finished the following four works.Firstly,carrying on the brief introduction to agent technology,automatic negotiation theory and negotiation strategy related method,summarizing and analyzing the research of the existing automatic negotiation system,and pointing out the existing problems,indicating the direction of improvement.Secondly,based on the reinforcement learning negotiation algorithms,proposing an innovative negotiating mechanism of opponent classification adaptive learning.The type of opponents is learned according to historical concession information,then the appropriate negotiation strategy is selected,and a large number of simulation experiments are performed to prove the effectiveness of the method.Thirdly,by analyzing the characteristics of people's performance in negotiations,taking into account that some cheating behaviors of human will appear in the negotiations,combining with the knowledge of psychology,a Human-Agent negotiation model with learning ability is constructed and a Human-Agent negotiation system is developed based on the model.Fourth,in order to verify whether the constructed adaptive learning method of opponent classification has excellent performance in Human-Agent negotiation,two comparative experiments are designed in this thesis.The experimental results show that the negotiation method of adding opponent classification strategy in Human-Agent negotiation can make the agent acquiring higher profits and improving the economic utility of Agent in Human-Agent negotiations.In this thesis,the reinforcement learning algorithm is applied to the Human-Agent negotiation system,and a learning method that recognizes the opponent's type is introduced.In terms of the research methodology,this thesis combines model construction,simulation experiment and empirical research,through which proves that the proposed model can make the automated negotiation system more effective.It is in line with the interests of the users represented by the Agent in the e-commerce environment.