Value-Centric Trust Model with Improved Familiarity Measurement

Jie Zhang, Ali A. GhorbaniIJCAI 2005 · OpenReview

Abstract

Formalization of familiarity contributes significantly to the formalization of trust within value-centric models. Traditional models often assume familiarity to be a static similarity of values between agents, which frequently results in low model stability. This paper proposes an improved familiarity measurement (TMIFM) that incorporates dynamic human-like factors such as interaction frequency, depth of engagement, and the passage of time (forgetting). Experimental results demonstrate that the TMIFM increases trust model stability by 33.47% compared to models using fixed familiarity values (TMFFV).

Introduction

In Multi-Agent Systems (MAS), establishing trust is critical for reliable interactions. While "value-centric" trust models have been developed, they often rely on a static interpretation of familiarity. In real-world scenarios, familiarity is not a fixed trait but a relationship that evolves through experience. Static familiarity leads to high variance in agent rankings, making it difficult for agents to make consistent long-term decisions. This research introduces a mathematical model for familiarity that mimics human cognitive processes to enhance system stability.

Methodology: The TMIFM Model

The proposed Trust Model with Improved Familiarity Measurement (TMIFM) treats familiarity as a dynamic variable updated after each transaction. The core of the model is a knowledge update equation:

$K_{current} = K_{previous} + \text{Processing Level} - \text{Forgetting Value}$

Key factors integrated into the model include:

  • Repeated Exposure: Familiarity increases with the frequency of transactions.
  • Level of Processing: Interactions involving higher workloads (more "widgets") contribute more to familiarity than minor ones.
  • Forgetting Rate: Familiarity decays over time based on the interval between interactions.
  • Sigmoid Mapping: Raw knowledge scores are mapped to a familiarity value between -1 and 1 using a logistic function to ensure computational stability.

Experimental Evaluation

The TMIFM was evaluated against a baseline model, the Trust Model with Fixed Familiarity Value (TMFFV), in a simulated agent society. The primary metric for success was the stability of agent rankings over time.

Figure 1: Comparison of agent rank stability over time between TMIFM and TMFFV.

As shown in Figure 1, the TMIFM (blue diamonds) exhibits significantly less "jitter" and reaches a stable state much faster than the TMFFV (pink squares). While the fixed model took approximately 40 days to stabilize, the improved model reached a stable average rank by day 15. This suggests that dynamic familiarity helps the system "push" agents to their correct rank more efficiently.

Results and Discussion

Quantitatively, the TMIFM demonstrated a substantial reduction in variance across multiple test runs. Stability in this context is defined by the consistency of an agent's rank; a lower variance indicates a more stable reputation system.

Figure 2: Quantitative comparison of stability (variance) across ten test runs.

The data in Figure 2 shows that the average variance for the TMIFM was 4.73, compared to 7.11 for the TMFFV. This represents a 33.47% improvement in stability. By mathematically modeling the "forgetting" and "processing depth" of interactions, the digital agents behaved more like a robust human community. This suggests that human-centric heuristics can effectively solve technical stability issues in distributed systems.

Conclusion

By incorporating cognitive realism into digital agent interactions, the TMIFM provides a more stable and reliable trust framework. The model successfully reduces ranking fluctuations, allowing for more predictable long-term agent behavior. Future work will focus on tuning the model's sensitivity parameters and testing its performance in noisy, real-world environments where malicious actors may attempt to manipulate reputation systems.

Summary updated

Abstract

Formalization of familiarity contributes to formalization of trust through a value-centric trust model. However, familiarity was assumed to be the similarity of values (fixed for two agents), and stability of the trust model was relatively low. To increase the stability, we propose an improved familiarity measurement. Experiments are carried out to compare the stability of the trust model with the improved familiarity measurement and with the fixed familiarity value. It is observed that the stability is increased by 33.47% through the improved familiarity measurement.
BibTeX
@inproceedings{zhang2005value,
  author = {Jie Zhang and Ali A. Ghorbani},
  title = {{Value-Centric Trust Model with Improved Familiarity Measurement}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2005},
  url = {https://openreview.net/forum?id=H1ZzjSfuWB},
}
Figure