A COMBINATION TRUST MODEL FOR MULTI-AGENT SYSTEMS

Abstract

Trust plays a crucial role in supporting agents to select the partners while interacting with each other in open distributed multi-agent systems. Researches on com- putational trust mostly focus on considering three types of trust: individual experience, inference trust from other agents in the community, and hybrid trust. Most current hy- brid trust models are the combination of experience and inference trusts and make use of some propagation mechanism to enable agents to share his/hernal trust with partners. However, the disadvantage of these models is that their use of such mechanism could result in high cost in computation. The problem with such a propagation based sharing may be reduced to the issue of considering which type of trusts an agent should utilize in evaluating partners and which type of trusts should be shared among agents. In this paper, werst propose a computational trust model which is a combination of experience and reference trusts. Our model may enable agents to evaluate the degree of trust on their partners, which is computed with a weighted combination based on the linguistic quantier function of their own experiences and the trustworthiness on their partners in the agent society. Then, we investigate which types of trust should be shared among agents and how they affect the effectiveness of interaction in our model. Our experiments are performed in an e-commerce environment and the experimental results have demon- strated that: (i) Changing two internal parameters, which are the linguistic quantier functions and weight vectors for aggregating the experience and reference trusts, may contribute to selecting the best partner for an agent but not much; (ii) combining the experience and inference trusts is better than using the single experience trust for esti- mating the trustworthiness of a partner; (iii) agents may share their single experience trust or combination trust with their partners and furthermore their making use of any of these sharing strategies may not affect the evaluation of product quality.