Research Of Automated Negotiation Based On Reinforcement Learning

With the development of computer technology, electronic commerce has come intoour lives. Traditional e-commerce just can be used to simulate simple scenes of life andis too simple to take advantage of negotiation. When people trade with others bytraditional e-commerce, they have to select agreement or refusal one-sided. But in reallife, people always use negotiation to resolve some of the differences and conflicts,which can enhance the understanding of buyers and sellers. The research of agentbecomes more and more mature with the aid of development of artificial intelligence.Because of growing ability of handling problem, automated negotiation combining withthe Agent technology and e-commerce can provide some solution for theabove-mentioned problems. Importing machine learning to automated negotiation is thecurrent mainstream research direction. The thesis researches the application ofreinforcement learning algorithm in negotiation.Firstly the thesis has a brief introduction of e-commerce, auto-negotiation theoryand the application of the common machine learning methods for auto-negotiation, thenresearches the traditional reinforcement learning algorithm, points out its problems, andproposes expectation restoration rate to improve original algorithm, at last appliesproposed negotiation strategy to bilateral multi-issue negotiation and verifies the effectof the improved algorithm mentioned by experimental comparison.Specific research work follows:â‘ the thesis researches the traditional reinforcement learning parameters (timediscount rate, time faith and so on), studies effect for negotiation when setting theparameters different values, points out how to select the appropriate parameters in thedifferent negotiation situations.â‘¡the thesis finds that the traditional reinforcement learning algorithm has theshortcoming of compromising too fast through experiment, Aiming at this problem, thisthesis proposes on improved reinforcement learning algorithm based on expectationrestoration rate to improve the original algorithm, and discusses the value of expectationrestoration rate. When expectation restoration rate is1, proposed algorithm becomesreinforcement learning algorithm.â‘¢Negotiation combined with opponent classification and reinforcement learning.By studying the history of the negotiation of the opponent, the opponent is divided into different categories, and each category has different negotiation attitude. When facingdifferent attitude negotiation opponent, agent has a different negotiation strategy.â‘£With the bilateral multi-issue negotiation framework, improved reinforcementlearning algorithm based on expectation restoration rate is applied to bilateralmulti-issue negotiation. Comparing with traditional reinforcement learning negotiationstrategy and time negotiation strategy through experiment results, proposed negotiationstrategy can get higher bilateral utility within allowing negotiation turns.