Agent Reputation and Trust models, formalisms and testbed --------------------------------------------------------- The seminar will be mainly composed of three parts. The first part presents the general motivations and problems tackled by trust and reputation models. The second part is dedicated to the formalisms and frameworks used to represent and reason about trust. The third part focuses on reputation by covering existing works and presenting the ART (Agent Reputation and Trust) testbed. C1. Motivations and Definitions for Trust and Reputation in Multi-Agent Systems The seminar begins with an introduction of the general motivations of trust in multi-agent systems. The typical problems tackled by trust and reputation model are presented. We show that they are different problems that the ones tackled by security techniques and that security and trust must be considered as complementary approaches. The foundations of trust models are then explained. Starting from sociological studies, we identify the main concepts involved in trust models. Clear definitions of concepts such as trust, reputation, recommendation, ... are given using the typology proposed by L. Mui (2002) and the functional ontology of reputation proposed by S. Casare (2005). C2. Trust Formalisms and Frameworks A survey of the main trust models in multi-agent systems is presented. The focus is both on conceptual and semi-formal models of trust (e.g. Castelfranchi & Falcone's model) and formal models of trust (e.g. Liau's model). The survey starts from formal models of informational trust, i.e. trust in information sources (Liau, Demolombe, Dastani et al.). In these models a certain agent is said to trust another agent if the former agent believes what the other agent says or that the information communicated to him by the other agent is reliable. The general model of trust proposed by Castelfranchi and Falcone is then presented. In Castelfranchi & Falcone's model, trust is defined as the trustor's evaluation of certain relevant properties of the trustee with respect to the achievement of a goal of the trustee. Recent logical formalizations and refinements of Castelfranchi & Falcone's model proposed by Lorini et al. are discussed. Jones's trust model is also presented in which trust is characterized in terms of two beliefs of the truster: the truster's belief that a certain rule or regularity applies to the trustee (called "rule belief"), and the truster's belief that the rule or regularity is going to be followed by the trustee (called "conformity belief"). The last part of the survey is devoted to present the logical approach to graded trust based on probability theory proposed by Josang. C3. Computational Reputation Models A survey of the main reputation models is presented. This survey starts from simple models based on a central repository of feedbacks as the ones used in e-commerce web site such as e-bay or OnSale and the Sporas and Histos models (Zacharia et al.). Reputation models that include an automatic decision process are then described (models from AbdulRahman et al. and from Marsh). At last, the most elaborated models that implement a complete automatic evaluation and reasoning process are presented: the models proposed by Sen et al., by Schillo et al., by Wang et al., the Liar model proposed by Muller et al., the Regret model proposed by Sabater and Sierra and the Repage model proposed by Paolucci et al.. At last, a presentation and a demonstration on the ART testbed (Fullam et al.) is done. The problem of the wide heterogeneity of reputation models is stated and the ART testbed is presented as a tool to allow the experimentation and the comparison of different models. The architecture and the functioning of ART are described. The seminar is concluded by a summary of the wide state of the art that has been presented and by giving a global prospective point of view of the tutors about the perspectives and the current challenges of the field.