The Semantic Interpretation of Trust in Multiagent Interactions

Anup K. KaliaAAAI 2014 · DOI

Introduction

In multiagent systems (MAS), establishing trust is a critical prerequisite for cooperation. Traditional computational models of trust often rely on numerical heuristics, such as interaction frequency or linguistic similarity. However, these methods frequently overlook the underlying semantics of interactions—specifically, the vulnerability of the truster and the autonomy of the trustee. This research by Anup K. Kalia proposes a shift toward a "semantic" interpretation of trust based on formal commitments, bridging the gap between sociological theory and practical algorithmic implementation.

The Commitment-Based Trust Model

The core of this approach is the use of commitments to represent formalized expectations between agents. A commitment is defined as $C(debtor, creditor, antecedent, consequent)$, where a debtor is obligated to fulfill a consequent if an antecedent condition is met. By tracking the lifecycle of these commitments—from creation to satisfaction (discharge) or violation—the system can derive a more nuanced understanding of agent behavior.

Interaction sequence between agents Buck and Selia illustrating commitment lifecycle and trust updates.

As shown in the figure above, trust is not a static value but a dynamic one that responds to specific outcomes. When a commitment is discharged (e.g., Buck pays Selia after delivery), trust increases. Conversely, if a commitment is canceled or violated, trust decreases.

To quantify this, the model adopts an evidence-based framework where trust is represented as a tuple $\langle r, s \rangle$, where $r$ denotes positive experiences (satisfied commitments) and $s$ denotes negative experiences (violated commitments). The probability of a positive outcome is calculated as $\alpha = \frac{r}{r+s}$. The model further personalizes this by introducing five parameters for each truster: initial bias ($r_{in}, s_{in}$), weights for satisfaction or violation ($i_r, i_s$), and a parameter for neutral interactions ($\lambda$).

Experimental Setup and Datasets

The model was evaluated using a combination of real-world and synthetic data to ensure a robust representation of commitment outcomes:

  • Enron Email Corpus: 33 emails were selected based on commitment-related language.
  • Synthetic Data: 28 synthetic sentences were added to balance the dataset with rarer commitment outcomes.
  • Human Judgments: 30 computer science students provided ground-truth trust values (0 to 1) after reading the interactions, which were used to train the model's parameters.

Key Results and Insights

The research yielded several significant findings:

  1. Predictive Accuracy: The model successfully learned user-specific parameters, minimizing the Mean Absolute Error (MAE) between predicted and human-reported trust values.
  2. Semantic Weighting: Results indicated that trust is driven more by the fulfillment of obligations than by interaction frequency. A single violation can outweigh multiple minor successes.
  3. Subjectivity: By adjusting $i_r$ and $i_s$, the model accurately reflected different human "forgiveness" thresholds—some users lose trust immediately upon a single failure, while others require a pattern of neglect.
  4. Contextual Impact: In game-based evaluations (using a variant of "Colored Trails"), the outcome of a specific commitment (e.g., receiving a promised tile) had a significantly higher impact on trust than the overall outcome of the game (winning or losing).

Conclusion and Future Work

This commitment-based approach provides a more human-like interpretation of trust in MAS by focusing on the semantics of promises. While the initial results are promising, future work will focus on scaling the evaluation using platforms like Amazon Mechanical Turk and integrating more complex linguistic features to capture the nuances of natural language. Additionally, the author aims to explore the recursive relationship between trust, mood, and emotion in virtual agent interactions.

Summary updated

BibTeX
@inproceedings{kalia2014semantic,
  author = {Anup K. Kalia},
  title = {{The Semantic Interpretation of Trust in Multiagent Interactions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8777},
}
Figure