With Jie Zhang, she explored scenarios where buying agents make use of advice provided by other buying agents in order to select the most appropriate selling agents. Included in this research is an approach that allows the combination of private and public modelling of the advisor's trustworthiness. The approach introduces as well incentives for honest reporting from advisors, through rewards provided by sellers to those advisors accepted into a large number of social networks of other buying agents.
With Reid Kerr, Professor Cohen has investigated the promotion of secure electronic marketplaces. This research first explored some of the common vulnerabilities in existing trust and reputation models, leading to a demonstration of how smart cheating agents can prosper even when their trustworthiness is being modeled. This in turn resulted in the design of a valuable testbed for measuring the performance of any trust and reputation modeling system. Most recently, work with Kerr focused on the critical challenge of collusion in multiagent systems, including techniques aimed at recognizing clusters of agents that avoid harm and instead produce benefit to each other.
In dynamic environments with sparse connections, it is beneficial to employ a multi-dimensional trust model, which takes into consideration the role of each agent that provides information, as well as its location, the time of the communication, its known trustworthiness from direct experience and its trustworthiness reflected from majority opinion in the multiagent system. This approach was explored for the application of mobile vehicular ad-hoc neworks, for which we developed as well an extensive simulation framework for trust-based travel decision making.
In environments where student learning can be achieved on the basis of the previous experience of peers, the challenge is to determine the appropriate social network to employ -- which peers are most reputable and which learning experiences have been most beneficial, for students bearing some similarity to the current student. This is the topic of research conducted with John Champaign, which also integrated a novel approach for reasoning about which commentary on learning objects to present to users, based on a modeling of reputation and similarity.
With Noel Sardana, research has investigated how to use trust in order to streamline the presentation of messages in social networking environments. An initial model integrated the concept of credibility, to balance consideration of similarity, to cope with folklore. From here, the research explored a principled Bayesian approach in order to learn from observations, against the backdrop of user utilities. This enables an examination of the use of trust modeling, for decision making. a
Most recently, Professor Cohen has been developing an approach to engender trust in multiagent systems, working with Thomas Tran. This enables a more complete viewpoint of the dynamics of trust modeling, examined from the perspective of the agent who is being modeled.
Other research has focused on the topic of multiagent resource allocation with preemption, conducted with John Doucette and Graham Pinhey, for cooperative environments. The challenge is to enable effective coordination of agents, even with dynamic task arrivals, through a modeling and exchange of plans and their utilities.
A longstanding instructor of social implications of computing, Professor Cohen has also recently been examining how to develop technological solutions to social problems of computers, including offering a graduate level course in this topic area.