A Distributed Reputation Mechanism Based on Multi-Agent Theories

In a centralized reputation system,there is an independent third party that calculates agents' reputation based on the feedback and transaction information,and then provides the reputation information for all agents.However,in a distributed system no independent third party exists.Transaction results and feedback information are stored in each agent.Thus,reputation information acquisition and aggregation become more complex.The main problems in previous studies are insufficiency of reputation information acquisition and inaccurate aggregation of reputation information from different sources. This paper proposes a computational model of reputation ina distributed reputation mechanism based on Multi-Agent System(MAS) theory.In distributed MAS,a trustor can get direct or indirect reputation information about a trustee by(1) relying on its direct transaction experience with the trustee;(2) finding witnesses who trade with the trustee through the trustor's social network;and(3) asking the trustee to refer its trading partners to the trustor as introducers.A model acquiring reputation information from these three sources is proposed.An incentive mechanism is included in the model,which is designed to encourage the trustee to honestly provide introducers' information in order to ensure the quantity and quality of acquired reputation information. Based on acquired information,a reputation information aggregation model is established to calculate the trustee's reputation.Because the reliability of reputation information varies with different sources,we propose a model to calculate the reliability of a trusty witness's evaluation based on the user similarity,and the reliability of an unknown introducer's evaluation.We further use the reliability test to weight corresponding reputation information.In addition,we introduce the default reputation for each role in each group in MAS,which is domain dependent knowledge and is applied for the first transaction of new entrants. By conducting experimental simulations,wecompare and analyze the effectiveness of our model and previous reputation models in different scenarios.The results provide evidence for strong validity and robustness of our model.It not only ensures a trustor to acquire much enough reputation information,but also effectively weakens the influence of lies and the bad intention of witnesses or introducers.The research conclusions contribute to current literatures on the sources of reputation information and improve the accuracy of evaluation outcomes in the distributed reputation mechanism.