Engineering Socially-Oriented Autonomous Agents and Multiagent Systems

Nieves MontesIJCAI 2022 · DOI

TL;DR

This paper proposes a two-pronged approach to building AI agents that can effectively interact within social contexts. It tackles both the societal level, by engineering norms and values, and the individual level, by equipping agents with "Theory of Mind" capabilities to understand others' perspectives. This research is crucial for developing more sophisticated and collaborative AI systems.

What Problem Does This Paper Address?

As AI systems become more integrated into our lives, there's a growing need for them to operate not just as individual tools, but as members of a community, capable of social interaction. This paper addresses the challenge of engineering "socially-aware" autonomous agents and multiagent systems. This involves two key aspects:

  1. Societal Level: How can we design systems of agents that adhere to shared rules and values, ensuring harmonious collective behavior? This is akin to creating a functional society for AI.
  2. Individual Level: How can we equip individual agents with the cognitive abilities to understand and predict the behavior of others, especially in cooperative or competitive scenarios? This involves developing a form of "Theory of Mind" for AI.

The paper aims to bridge the gap between purely functional AI and AI that can navigate the complexities of social dynamics.

Methods

The research is divided into two main approaches:

  • Societal Level (MAS Level):
    • Value-Guided Norm Synthesis: A methodology for automatically generating prescriptive norms that align with specific societal values. Norms are represented as optimizable parameters, allowing the use of metaheuristic techniques to find norm sets that best promote desired values. This includes tools to analyze the contribution of individual norms (Shapley values) and the compatibility between different values.
  • Action Situation Language (ASL): A novel language inspired by Elinor Ostrom's work, designed to represent a wide variety of norms in a machine-readable format (if-then-where statements). ASL enables agents to perform "what-if" analyses on the impact of adopting norms. It's complemented by a game engine that formalizes these situations as extensive-form games, allowing for standard game-theoretical analysis. A key feature of ASL is its extensibility, allowing for automated assessment of changes to rules and their priorities.
  • Individual Level:
    • Theory of Mind (ToM) + Abductive Reasoning: This approach focuses on the cognitive machinery an agent needs to reason about others. It involves an observer agent simulating another agent's perspective (which is often incomplete) and then using abductive reasoning to infer the explanations that would justify the observed action. These inferred explanations are then incorporated into the observer's knowledge base for future decision-making.

Datasets

Not discussed in the paper. The paper focuses on developing methodologies and frameworks, with the game of Hanabi mentioned as a testbed for evaluating the individual-level approach.

Key Insights and Ideas

  • Norms as Optimizable Parameters: Representing norms as tunable parameters, rather than fixed rules, opens up powerful avenues for automatically synthesizing normative systems that effectively promote desired societal values.
  • ASL for Social Rule Analysis: The Action Situation Language provides a flexible and formal way to model and analyze social rules and their impact, making it a valuable tool for understanding complex social interactions in AI.
  • Theory of Mind for Cooperation: Equipping agents with Theory of Mind is not just for competitive scenarios; it can significantly boost performance in cooperative tasks by enabling better understanding of teammates' intentions and perspectives.
  • Abduction for Inferring Intent: Using abductive reasoning to infer the "why" behind an agent's action is a powerful way to enrich an agent's understanding of others, especially in situations with imperfect information.
  • Domain-Agnostic Cognitive Machinery: The core cognitive mechanisms for social awareness (like ToM + abduction) are designed to be independent of the specific domain, suggesting broad applicability.

Key Results

The paper presents methodologies and frameworks rather than specific quantitative results from empirical evaluations.

  • The value-guided norm synthesis methodology allows for the automatic generation of normative systems that promote specific values.
  • The ASL framework has been used to model benchmark social scenarios from policy analysis literature.
  • The ToM + abduction approach is being explored as a way to improve agent performance in cooperative tasks, with the game of Hanabi serving as a primary testbed.

Limitations and Future Work

  • Limitations of Norm Representation: The initial norm synthesis method had a rigid representation of norms, requiring redefinition for different scenarios. ASL was developed to address this.
  • Future Work:
    • Extending ToM: Moving beyond first-order ToM to allow for more complex epistemic strategies.
  • Strategy Optimization: Automatically searching for optimal strategies within the ASL framework, aligning with the optimization approach from the societal level research.
  • Ad Hoc Teamwork: Extending the ToM+abduction machinery for situations where agents must coordinate without a pre-agreed strategy, a significant challenge in multiagent systems.

Summary updated

BibTeX
@inproceedings{montes2022engineering,
  author = {Nieves Montes},
  title = {{Engineering Socially-Oriented Autonomous Agents and Multiagent Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/833},
}
Figure