In the fall of 2013 I did an undergraduate research assistantship with professors Robin Cohen and Thomas Tran on the topic of earning trust in multi-agent systems. A multi-agent system consists of many interacting autonomous agents in an environment. Such systems are common in the real world: traffic, marketplaces, and computer networks are all examples. The study of systems in which many independent agents act to further their own goals helps us understand and improve these real-world applications.
Trust is an important concept in multi-agent systems. Trust is an expectation of how other agents will behave, and establishing trustworthiness allows an agent to make better decisions with less risk. In the paper A Comprehensive Approach to Trust Management, Sandip Sen discusses the notion of trust in multi-agent systems and notes that although the problem of evaluating the trustworthiness of others has
been well studied, there has been less interest in the complementary problem of earning trust. My research was based on this paper and Sen's call for study into trust establishment, among other components of a proposed comprehensive trust management scheme.
My research involved simulating and studying trust establishment in the context of an electronic marketplace. The simulated environment consisted of buyers (trustors) and sellers (trustees). The seller provides a service at some price and quality. If the buyer is satisfied with the price and quality then its trust in the seller increases, otherwise it decreases. I implemented an exploratory simulation filled with several categories of buyers and sellers. The purpose of this simulation was to provide a baseline understanding of some simple trustee behaviours and to evaluate the effectiveness of the simulation itself.
The exploratory simulation contained a population of three types of buyers: price-sensitive, balanced, and quality-sensitive; indicating the relative value of low price vs. high quality transactions. The seller population consisted of profit-maximizing sellers that always offered the same price and quality, randomized sellers that randomly picked some price and quality at fixed profit, and classifying sellers that attempted
to predict the buyer category then made a corresponding offer.
The simulation results demonstrated that the sellers which predicted the buyer's desires were more effective at earning trust than prot-maximizing or randomized sellers. However, the environment proved too simple to evaluate less trivial trustee strategies. The sellers were too easy to predict and consistently satisfy.
My work resulted in software for simulating an electronic marketplace environment with a variety of buyer and seller agents. Prediction of trustor desires was shown to be a promising strategy for earning trust. Finally a list of recommendations was created for improving the sellers and the simulation environment to permit testing of more interesting trust-earning behaviours.