Designing and trusting multi-agent systems for B2B applications

TLDR: A trust model allowing agents to evaluate the credibility of other peers in the environment using agents' credibility is proposed, which applies a number of measurements in trust evaluation of other party's likely behavior.

Abstract: This thesis includes two main contributions. The first one is designing and implementing B usiness-to-B usiness (B2B ) applications using multi-agent systems and computational argumentation theory. The second one is trust management in such multi-agent systems using agents' credibility. Our first contribution presents a framework for modeling and deploying B2B applications, with autonomous agents exposing the individual components that implement these applications. This framework consists of three levels identified by strategic, application, and resource, with focus here on the first two levels. The strategic level is about the common vision that independent businesses define as part of their decision of partnership. The application level is about the business processes, which are virtually integrated as result of this common vision. Since conflicts are bound to arise among the independent applications/agents, the framework uses a formal model based upon computational argumentation theory through a persuasion protocol to detect and resolve these conflicts. Termination, soundness, and completeness properties of this protocol are presented. Distributed and centralized coordination strategies are also supported in this framework, which is illustrated with an online purchasing case study followed by its implementation in Jadex, a java-based platform for multi-agent systems. An important issue in such open multi-agent systems is how much agents trust each other. Considering the size of these systems, agents that are service providers or customers in a B2B setting cannot avoid interacting with others that are unknown or partially known regarding to some past experience. Due to the fact that agents are self-interested, they may jeopardize the mutual trust by not performing the actions as they are supposed to. To this end, our second contribution is proposing a trust model allowing agents to evaluate the credibility of other peers in the environment. Our multi-factor model applies a number of measurements in trust evaluation of other party's likely behavior. After a period of time, the actual performance of the testimony agent is compared against the information provided by interfering agents. This comparison process leads to both adjusting the credibility of the contributing agents in trust evaluation and improving the system trust evaluation by minimizing the estimation error.