Introducing trust and reputation into multi-agent systems can significantly improve the quality and efficiency of the systems. The computational trust and reputation also creates an environment of survival of the fittest to help agents recognize and eliminate malevolent agents in the virtual society. The research redefines the computational trust and analyzes its features from different aspects. A systematic model called Neural Trust Model for Multi-agent Systems is proposed to support trust learning, trust estimating, reputation generation, and reputation propagation. In this model, the research innovates the traditional Self Organizing Map (SOM) and creates a SOM based Trust Learning (STL) algorithm and SOM based Trust Estimation (STE) algorithm. The STL algorithm solves the problem of learning trust from agents' past interactions and the STE solve the problem of estimating the trustworthiness with the help of the previous patterns. The research also proposes a multi-agent reputation mechanism for generating and propagating the reputations. The mechanism exploits the patterns learned from STL algorithm and generates the reputation of the specific agent. Three propagation methods are also designed as part of the mechanism to guide path selection of the reputation. For evaluation, the research designs and implements a test bed to evaluate the model in a simulated electronic commerce scenario. The proposed model is compared with a traditional arithmetic based trust model and it is also compared to itself in situations where there is no reputation mechanism. The results state that the model can significantly improve the quality and efficacy of the test bed based scenario. Some design considerations and rationale behind the algorithms are also discussed based on the results.
The Packet-World: A Test Bed for Investigating Situated Multi-Agent Systems