Adaptable and Verifiable BDI Reasoning

Peter Stringer, Rafael C. Cardoso, Xiaowei Huang, Louise A. DennisarXiv 2020 · DOI

TL;DR

This paper proposes a framework for autonomous systems to adapt to changes in their environment by developing an agent-maintained "self-model." This model tracks action performance, detects failures, and enables learning new action descriptions to maintain long-term autonomy, crucial for complex systems like Mars rovers.

What Problem Does This Paper Address?

Autonomous systems, especially those operating for extended periods in dynamic environments (like Mars rovers or satellites), need to adapt when their capabilities degrade or fail. Traditional Belief-Desire-Intention (BDI) systems, while good for intelligent agent programming, often lack robust mechanisms for detecting and recovering from action failures. This paper addresses the challenge of creating adaptable BDI agents that can learn from experience, update their understanding of the world and their own actions, and reconfigure their plans to maintain functionality in the face of unexpected changes or performance degradation.

Methods

The core of the proposed solution is an agent-maintained self-model, acting as a formal ontology for the autonomous system. This self-model stores information about the system's processes and functionality, including action descriptions with explicit pre- and post-conditions, inspired by the Planning Domain Definition Language (PDDL).

Key methodological contributions include:

  • Durative Actions: Introducing a theory of durative actions for BDI systems, allowing agents to continue operating while an action is in progress and monitor its status. This moves beyond the typical BDI approach where an agent is suspended until an action completes.
  • Action Failure Detection: Developing a mechanism to monitor an action's life-cycle. Actions can transition from "Functional" to "Suspect" and then "Deprecated" based on persistent abnormal behavior, detected by comparing actual post-conditions with expected ones.
  • Learning New Action Descriptions: Integrating machine learning techniques to learn new action descriptions when existing ones fail. This allows the system to adapt to novel situations or environmental changes.
  • Plan Reconfiguration: Leveraging AI planners to reconfigure existing plans by replacing failing actions with alternative ones or newly learned actions, ensuring the system can still achieve its goals.
  • Verification: Utilizing the GWENDOLEN agent programming language and the MCAPL framework with the AJPF model-checker to verify the adapted system and ensure safety properties are maintained.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • The Self-Model as a Central Hub: The self-model is not just a passive repository but an active component that integrates information from various sources (action execution, learning, planning) to provide a coherent understanding of the agent's capabilities and their performance.
  • Proactive Failure Management: Instead of just reacting to complete failures, the system aims to detect degradation early (moving to a "Suspect" state) and take corrective action before a complete failure occurs.
  • Bridging Planning and BDI: The work effectively bridges classical AI planning (PDDL) with BDI agent architectures, leveraging the strengths of both for robust autonomous behavior.
  • Learning for Adaptability: Machine learning is presented not just as a way to improve performance but as a fundamental mechanism for enabling long-term adaptability in dynamic and unpredictable environments.
  • Verification is Key for Safety: The emphasis on integrating with model-checking tools like AJPF highlights the critical need to ensure that adaptation and learning do not compromise the safety and reliability of autonomous systems.

Key Results

This paper is a position paper outlining a research program and proposing a system architecture. Therefore, it does not present empirical results from a deployed system. The key "results" are the conceptual framework and the proposed mechanisms for:

  • Representing actions with explicit pre- and post-conditions, including durative aspects.
  • Monitoring action life-cycles to detect failures.
  • Learning new action descriptions.
  • Reconfiguring plans using AI planners.
  • Integrating these components within a verifiable BDI framework.

Limitations and Future Work

The authors acknowledge several limitations and outline future work:

  • Subjectivity of "Persistent Failure": The term "persistent failure" needs a formal and precise specification to avoid ambiguity.
  • Post-Reconfiguration Behavior: Further work is needed to define what happens to failing actions in the model after reconfiguration.
  • Learning Strategy Challenges: The proposed learning strategy presents challenges, including ensuring valid solutions, minimizing planning time, and how an action's state influences the learning strategy.
  • Formal Definition of Self-Model: A formal definition of the self-model and its included concepts is required.
  • Implementation and Evaluation: The proposed system architecture needs to be implemented and evaluated.
  • Reverification Necessity: Further study is needed to identify conditions where reverification of the entire system might not be necessary after a new action is learned.
  • Safe Learning: The risk of system properties being violated during machine learning is acknowledged, suggesting the need for "Safe Learning" approaches, especially for safety-critical applications like autonomous drones.

Summary updated

BibTeX
@misc{stringer2020adaptable,
  author = {Peter Stringer and Rafael C. Cardoso and Xiaowei Huang and Louise A. Dennis},
  title = {{Adaptable and Verifiable BDI Reasoning}},
  year = {2020},
  doi = {10.4204/eptcs.319.9},
}
Figure