A decentralized trustworthiness estimation model for open, multiagent systems (DTMAS)

Abstract

Often in open multiagent systems, agents interact with other agents to meet their own goals. Trust is, therefore, considered essential to make such interactions effective. However, trust is a complex, multifaceted concept and includes more than just evaluating others' honesty. Many trust evaluation models have been proposed and implemented in different areas; most of them focused on algorithms for trusters to model the trustworthiness of trustees in order to make effective decisions about which trustees to select. For this purpose, many trust evaluation models use third party information sources such as witnesses, but slight consideration is paid for locating such third party information sources. Unlike most trust models, the proposed model defines a scalable way to locate a set of witnesses, and combines a suspension technique with reinforcement learning to improve the model responses to dynamic changes in the system. Simulation results indicate that the proposed model benefits trusters while demanding less message overhead.

The Social Concept of Trust as Enabler for Robustness in Open Self-Organising Systems

“…Due to the subjective nature of trust and because agents might lie about the trustworthiness of others, it is often also desirable to weigh the impact a recommending agent, called witness, has on the reputation value. The neighbour trust metric [OCT6] as well as DTMAS [28] propose to increase the influence of a witness with the similarity of the provided valuation to the one of the requesting agent. If the difference is too large, the witness can even be excluded from the calculation.…”

A hybrid trust model using reinforcement learning and fuzzy logic

“…A well‐known issue with RL‐based trust models is that trusters do not quickly recognize the environment changes and adapt to new settings. To address this shortcoming, DTMAS uses a technique similar to the regret described in the work of Marsh and Briggs to improve the responses of RL‐based models to dynamic changes in the system. The scholars proposed suspending the use of a trustee immediately as a response to an unsatisfactory transaction with the trustee.…”

“…Q‐learning does not depend on the use of complete and accurate knowledge to take proper actions, and therefore, agents can learn from their own experience to perform the best actions . Unfortunately, the optimization process can be sensitive to the selection of the reinforcement signal and the fact that the system states must be visited a sufficient number of times, which is alleviated by the use of temporal suspension as in the work of Aref and Tran …”

“…The idea is that accidentally misbehaving trustees keep their reputation in the community unlike those who misbehave due to a long‐term behavior change. Similarly, an agent may suspend the use of testimony from some witnesses …”

“…We define 3 input parameters for the main FLS as follows: The calculated direct trust (DT) using Q‐learning in a way similar to what was used in the work of Aref and Tran was based on the work of Tran and briefly presented here for convenience. This parameter represents the long‐term relationship between a truster and the trustee.…”