Combining Direct Trust and Indirect Trust in Multi-Agent Systems

Elham Parhizkar, Mohammad Hossein Nikravan, Robert C. Holte, Sandra Zilles

Abstract

In multi-agent systems (MAS), agents must navigate interactions with potentially dishonest peers by leveraging both personal experience (direct trust) and third-party advice (indirect trust). This paper presents a systematic study of the optimal integration strategies for these two trust sources. We evaluate four state-of-the-art indirect trust models across 14 distinct dishonest advisor attack scenarios. Our findings introduce ExpFun, a novel exponential decay strategy that dynamically weights advice based on the agent's growing personal experience. Experimental results demonstrate that ExpFun is the most robust integration method, particularly effective against dynamic attacks where traditional models falter.

Introduction

Trust modeling is essential for cooperation in decentralized multi-agent environments. Agents typically rely on Direct Trust, derived from their own interaction history, and Indirect Trust, synthesized from community ratings. However, both sources have inherent weaknesses: direct trust suffers from the "cold start" problem in new environments, while indirect trust is vulnerable to malicious advisors who engage in "ballot-stuffing" (inflating reputations) or "badmouthing" (deflating reputations).

Despite the proliferation of algorithms to filter dishonest advice, the optimal "mixing ratio" between personal experience and external advice has remained under-researched. This study investigates whether combining these sources is universally beneficial and identifies the most effective mathematical framework for doing so.

Methodology: The ExpFun Strategy

The researchers compared several integration strategies, including simple averaging, reinforcement learning (ACT-RL), and treating direct experience as just another advisor. The standout contribution is ExpFun, which utilizes an exponential decay function to determine the weight of indirect advice ($w_{ind}$):

$$w_{ind} = e^{-\lambda(p+n)}$$

In this model, $(p+n)$ represents the total number of direct interactions. As an agent gains more personal experience, the reliance on external advice decreases exponentially. This allows for a seamless transition from community-reliance during the initial stages to self-reliance as the agent's own data becomes statistically significant.

Evaluation and Results

The performance of these strategies was measured using the Relative Frequency of Unsuccessful interactions (RFU) across 14 attack scenarios, ranging from static badmouthing to complex dynamic attacks like "Camouflage" (where advisors act honestly to build trust before lying).

Table 1: RFU performance for various trust models and integration strategies under static and selective attacks.

As shown in Table 1, relying solely on indirect trust (OI) can be catastrophic in selective attack scenarios. For instance, in the "Selective Badmouthing" (BM) scenario with 80% dishonest advisors, the RFU for the ITEA model spikes to 0.551. However, the application of ExpFun (EF) consistently maintains lower RFU scores, often hovering between 0.14 and 0.15 in standard scenarios, demonstrating its ability to mitigate the impact of even widespread advisor dishonesty.

Dominance Analysis

To confirm the generalizability of these findings, a dominance analysis was conducted across all 14 experimental settings. This analysis tracked how often one integration method outperformed its peers.

Table 2: Dominance analysis comparing ExpFun against other integration strategies across 14 attack scenarios.

Table 2 highlights the clear superiority of the ExpFun method. In the "Camouflage" scenario, ExpFun achieved a dominance score of 12,0, meaning it outperformed other methods in every tested configuration without a single loss. This dominance is particularly pronounced in dynamic scenarios (Camouflage, Periodic, Good-Bad-Good), where an agent's recent personal experience is the only reliable indicator of shifting advisor behavior.

Conclusion

This research concludes that while strong indirect trust algorithms are powerful, they are not a complete solution. The integration of direct experience is vital, especially in dynamic environments. The ExpFun strategy provides a robust, mathematically simple, and highly effective way to balance these trust sources, outperforming complex reinforcement learning approaches and simple heuristics alike. Future work will focus on optimizing the sensitivity parameter ($\lambda$) and testing the scalability of these methods in larger, more complex agent networks.

Summary updated

Abstract

To assess the trustworthiness of an agent in a multi-agent system, one often combines two types of trust information: direct trust information derived from one's own interactions with that agent, and indirect trust information based on advice from other agents. This paper provides the first systematic study on when it is beneficial to combine these two types of trust as opposed to relying on only one of them. Our large-scale experimental study shows that strong methods for computing indirect trust make direct trust redundant in a surprisingly wide variety of scenarios. Further, a new method for the combination of the two trust types is proposed that, in the remaining scenarios, outperforms the ones known from the literature.
BibTeX
@inproceedings{parhizkar2020combining,
  author = {Elham Parhizkar and Mohammad Hossein Nikravan and Robert C. Holte and Sandra Zilles},
  title = {{Combining Direct Trust and Indirect Trust in Multi-Agent Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/44},
}
Figure