Deontic Sensors

Julian A. Padget, Marina De Vos, Charlie Ann PageIJCAI 2018 · DOI

TL;DR

This paper introduces "deontic sensors," a novel Resource-Oriented Architecture (ROA) pattern that decouples normative reasoning from multi-agent systems (MAS) platforms. This allows for more flexible integration of normative capabilities, enabling certification, reuse, and easier access to "on-demand" normative reasoning for open MAS.

What Problem Does This Paper Address?

Multi-agent systems (MAS) often need to incorporate normative reasoning – rules about what agents "ought" or "ought not" to do. Traditionally, this normative reasoning is either tightly coupled within specific MAS platforms, making it inaccessible and difficult to reuse, or it's entirely absent. This "siloed" approach hinders the development and sharing of normative reasoning components, limits transparency, and makes establishing trust challenging. The authors aim to break down these silos by making normative reasoning a discoverable and reusable service.

Methods

The core contribution is the "deontic sensor" pattern, framed within a Resource-Oriented Architecture (ROA). This pattern treats normative reasoning as a service that agents can query.

  • Deontic Sensors: These are conceptualized as transducers that take "brute facts" (agent actions) from the environment and interpret them into "institutional facts" (normative interpretations).
  • ROA Pattern: Instead of a rigid API, the pattern defines a set of Create, Read, Update, and Delete (CRUD) operations over resources (like institutional models and their instances). This allows for loose coupling, where the normative reasoning service (the institution) and the client (the agent platform) are agnostic to each other's internal implementation.
  • Decoupling: The architecture separates the agent platform (AP) from the deontic sensor platform (DSP). Agent actions are sent to the DSP for interpretation, and normative feedback is returned to the AP as "deontic commentary" (percepts) for agents.
  • Instantiation: The paper demonstrates an instantiation using the Jason agent platform and the InstAL (Institutional Action Language) for normative reasoning, exposed via a RESTful interface (InstAL-REST).

Datasets

Not discussed in the paper. The evaluation uses synthetic workloads to test performance.

Key Insights and Ideas

  • Normative Reasoning as a Service: Framing normative reasoning as a ROA service is a significant shift, enabling it to be treated like any other web service. This opens up possibilities for discovery, composition, and reuse.
  • Loose Coupling for Flexibility: The ROA pattern's emphasis on resource manipulation rather than strict APIs allows for greater flexibility. Agent platforms and normative frameworks can be swapped out or updated independently.
  • Enabling Trust and Certification: By making normative reasoning accessible as a service, it becomes possible to verify and certify its behavior, fostering trust in MAS.
  • Democratizing Normative Reasoning: The approach aims to make normative reasoning accessible to non-specialists, allowing them to query and understand the normative implications of actions on demand.

Key Results

The paper presents a conceptual pattern and a sample architecture. The evaluation focuses on the performance and scalability of their InstAL-REST instantiation.

  • Scalability: The system demonstrated scalability under simulated agent workloads, with response times increasing but not linearly with the number of agents or queries, indicating effective amortization of network overhead.
  • Performance: For batched queries, the average time per query decreased significantly as the batch size increased, showing that network overheads outweigh query costs.
  • Demonstration of Concept: The paper successfully demonstrates how an agent (Alice) can interact with multiple institutions, applying for permits and performing actions based on normative feedback, as illustrated in Listings 1 and 2.

Limitations and Future Work

  • Interoperability: While the pattern decouples components, achieving interoperability still relies on compatible percept conventions and data representations between the agent platform and the deontic sensor. Transcoding mechanisms might be needed.
  • Normative Percept Language: The paper acknowledges that the percept language for normative feedback can be flexible, but standardization or resolution mechanisms might be necessary for broader adoption.
  • Future Work: The authors suggest distributing deontic sensors across compute resources, applying them to agent-based simulations, and investigating semantic technologies to address the "semiotics" of deontic sensors.

Summary updated

BibTeX
@inproceedings{padget2018deontic,
  author = {Julian A. Padget and Marina De Vos and Charlie Ann Page},
  title = {{Deontic Sensors}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/66},
}

Loading related content...

Figure