Robust Norm Emergence by Revealing and Reasoning about Context: Socially Intelligent Agents for Enhancing Privacy

Nirav Ajmeri, Hui Guo, Pradeep K. Murukannaiah, Munindar P. SinghIJCAI 2018 · DOI

TL;DR

This paper introduces "Poros," a novel framework for socially intelligent agents (SIPAs) that learn and adapt to social norms by revealing and reasoning about contextual information during interactions. By allowing agents to share context when deviating from norms, Poros significantly improves social cohesion and enhances individual goal satisfaction compared to traditional approaches that rely solely on sanctions.

What Problem Does This Paper Address?

Social norms are the unwritten rules that govern human interactions, ensuring predictable and cooperative behavior. While existing computational approaches to norm emergence often simplify interactions to basic deviations and sanctions, real-world norm adherence is far more nuanced. Deviations from norms are not always negative; they can be justified by specific contexts. For instance, answering a phone call during a meeting is usually a violation, but doing so to address a family emergency might be acceptable. The challenge is to create artificial agents that can understand these contextual nuances, learn robust norms, and maintain social harmony while still allowing for individual autonomy and goal achievement. This paper addresses the gap in existing systems by proposing a method for agents to proactively share contextual information to explain deviations, leading to more sophisticated norm learning and better social outcomes.

Methods

The core of this paper is the Poros framework, which enables Socially Intelligent Personal Agents (SIPAs) to engage in "enriched interactions." Unlike simpler models where agents either comply or deviate and are then sanctioned, Poros agents can:

  • Reveal Context: When a SIPA deviates from a norm, it can choose to reveal specific elements of its context (e.g., the caller's relationship, the urgency of a call) to other agents. This is a key innovation, moving beyond simple deviation/sanction.
  • Reason about Context: Other agents, upon receiving revealed context, use this information to re-evaluate the deviation. They consider what they would have done in that context and adjust their sanctions accordingly.
  • Learn Norms: Through repeated interactions where context is revealed and reasoned about, agents collectively refine their understanding of norms. They learn which deviations are acceptable under specific circumstances, leading to more robust and contextually relevant norms.

The framework models agents as having goals and plans, and their interactions are governed by a social architecture comprising norms and sanctions. The "social experience" for an agent is defined as a weighted aggregation of payoffs for its stakeholders, allowing for different agent types (pragmatic, considerate, selfish) based on how they prioritize these payoffs.

Datasets

Not discussed in the paper. The experiments were conducted using a simulated "ringer environment" built with the MASON simulation framework. This environment was populated with agents representing different network types (large/small, dense/sparse) and agent societies (pragmatic, considerate, selfish).

Key Insights and Ideas

  • Context is King for Norms: The most significant insight is that simply sanctioning deviations is insufficient for robust norm learning. Revealing and reasoning about the context of a deviation is crucial for agents to understand the boundaries of norms and for those norms to become more sophisticated and adaptable.
  • Beyond Negative Sanctions: While negative sanctions are important for establishing norms, the ability to explain deviations with context can lead to positive reinforcement or at least a reduction in negative sanctions, fostering better social cohesion.
  • Improved Goal Satisfaction: By learning more nuanced norms, agents can better satisfy their individual goals. This is because they are less likely to be unfairly sanctioned for contextually justified deviations, and they can also better predict and navigate social interactions.
  • Agent Type Matters: The way agents aggregate payoffs (pragmatic, considerate, selfish) influences their behavior and the outcomes, but Poros consistently shows benefits across these types, highlighting the robustness of the context-revealing mechanism.

Key Results

The paper's experiments, conducted in a simulated "ringer environment" with various agent types and network structures, yielded strong results:

  • Social Cohesion: Poros agents achieved significantly higher social cohesion (82.0% - 83.2%) compared to both Fixed agents (around 52%) and Sanctioning agents (around 61-63%). This means fewer negative sanctions were issued.
  • Social Experience (Goal Satisfaction): Poros agents delivered substantially higher social experience (2.14 - 2.19) than Fixed agents (0.53 - 0.56) and Sanctioning agents (1.11 - 1.22). This indicates better goal satisfaction for the agents' users.
  • Robustness: These improvements were consistent across different network sizes and densities, and also held true, albeit with some variations, in considerate and selfish agent societies.

Limitations and Future Work

  • Simulation-Based: The findings are based on simulations. Real-world testing with end-users is needed to validate adaptability and user preferences.
  • Universal Context Revelation: In the current Poros implementation, agents always reveal context when deviating. The authors acknowledge that ideally, SIPAs should selectively reveal context to protect user privacy. This selective revelation is a key area for future work.
  • Eliciting User Attitudes: Reliably eliciting user attitudes and preferences for privacy and norm adherence remains a challenge.
  • Affect and White Lies: Future directions include incorporating affect into norm understanding and exploring the role of "white lies" (e.g., fabricating a mild excuse instead of a severe one) in promoting privacy and social cohesion.
  • Argumentation for Context Selection: Incorporating argumentation frameworks, similar to Kökciyan and Yolum's work, could help SIPAs decide which subset of context to reveal.

Summary updated

BibTeX
@inproceedings{ajmeri2018robust,
  author = {Nirav Ajmeri and Hui Guo and Pradeep K. Murukannaiah and Munindar P. Singh},
  title = {{Robust Norm Emergence by Revealing and Reasoning about Context: Socially Intelligent Agents for Enhancing Privacy}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/4},
}
Figure