TL;DR
This paper introduces an agent-based system for diagnosing and adapting models used in plan execution for complex domains like military air operations. By analyzing "expectation failures" – small discrepancies between planned and actual behavior – the system generates hypotheses about model errors, prioritizes them, and suggests repairs to human operators, aiming to improve planner adaptability and performance.
What Problem Does This Paper Address?
In domains like military air operations, complex plans are generated based on a vast set of hierarchical models describing entities and processes. When these plans execute, the actual behavior might deviate from expectations. While catastrophic failures are obvious, "non-critical anomalies" – minor deviations that still allow the goal to be achieved – are harder to detect and diagnose. These anomalies, however, are crucial for refining the underlying models and improving future plan performance. Manually diagnosing and fixing model errors in large-scale, multi-agent systems with thousands of concurrent plans is infeasible. This paper addresses the challenge of automating the detection, evaluation, and revision of these models based on subtle expectation failures during plan execution.
Methods
The core of the proposed solution is an agent-based diagnosis system integrated with a mixed-initiative model adaptation component.
- Agent-Based Diagnosis: The system employs specialized agents, each representing a hypothesis about a specific model failure. These agents gather data from historical plan executions and current anomalies to calculate the likelihood of their hypothesis being true.
- Hypothesis Generation: Hypotheses are generated through four primary methods:
- Initialization: A set of general hypotheses (e.g., old model version, non-model error) are always considered.
- Anomaly-Triggered: When a plan anomaly is detected, the system analyzes the context and related models to generate hypotheses about specific model classes and parameters.
- Goal Failure: If a plan fails to satisfy a goal, hypotheses are generated about inconsistencies between actor capabilities, planner reasoning models, and world state changes.
- Generalization/Refinement: As an agent's confidence grows, it can generalize hypotheses by traversing the model hierarchy or refine them by pinpointing specific parameters or model interactions.
- Hypothesis Refinement: This involves narrowing down the potential error to specific parameters or interactions between models. For example, if fuel anomalies are detected for an A-10 aircraft, the system might refine the hypothesis to suspect the
fuelBurnRateparameter for navigate activities. - Likelihood Calculation: Agents use various tasks, including querying model and historical plan case bases, analyzing current execution data, and performing statistical analysis, to compute the belief in their hypothesis. Dempster-Shafer theory and operator-modifiable likelihood functions are used for evidence combination.
- Mixed-Initiative Model Adaptation: The system presents the most confident hypotheses as suggestions for model repair to a human user. The user can accept, reject, or ignore these suggestions, providing feedback that helps the system refine its future suggestions.
Datasets
The paper uses three artificial datasets created to represent different types of expectation failures:
- Data Set 1: Represents a typical case where a fuel burn rate in the model is too high, leading to a predicted fuel level lower than actual.
- Data Set 2: Represents a scenario with false anomalies due to a faulty sensor reporting all fuel values as zero.
- Data Set 3: Represents a scenario where an aircraft runs out of fuel and crashes, with spatial anomalies also appearing.
The authors note that most of their real data is proprietary and thus not used for public evaluation.
Key Insights and Ideas
- Value of Non-Critical Anomalies: The paper highlights that even small, non-catastrophic deviations from plan expectations are valuable signals for model refinement. These "anomalies" are the primary basis for learning and improving the planner's background knowledge.
- Agent-Based Hypothesis Generation: Using individual agents for each hypothesis allows for a distributed and scalable approach to exploring a vast space of potential model errors.
- Model Hierarchy Leverage: The system effectively uses a hierarchy of models to generalize and refine hypotheses, moving from broad suspicions (e.g., "A-10 model error") to specific ones (e.g., "fuelBurnRate on A-10 for navigate activities").
- Mixed-Initiative is Crucial: While automation is key, the authors emphasize that a human-in-the-loop is essential for handling complex or ambiguous situations, especially when faults lie outside the modeled system or when visual cues from other tools are important.
- Temporal Link Analysis: This technique is vital for correctly attributing anomalies to specific activities by examining the sequence of events and the progression of discrepancies.
Key Results
The paper presents a qualitative comparison of their agent-based approach against a "naïve" hand-coded rule-based system.
- Data Set 1: Both methods achieved comparable results, correctly identifying a fuel burn rate parameter error for navigate activities on the A-10 model. The agent-based system's temporal link analysis was crucial for identifying the correct activity.
- Data Set 2: The agent-based system showed potential for handling false anomalies by incorporating a knowledge-based sanity check, though it still required human intervention to correctly interpret the situation.
- Data Set 3: This dataset, representing a crash scenario, was challenging for both the agent-based system and the rules alone, requiring user insight to correctly diagnose the fault.
Limitations and Future Work
- Human-in-the-Loop Necessity: The authors acknowledge that their system, even with mixed-initiative, cannot handle all classes of faulty anomaly reporting without human intervention.
- Problem of Focus: Users might ignore suggestions due to other pressing issues or conflicting visual information. Future work needs to address how to better guide user attention.
- Handling External Faults: The system struggles with faults outside the modeled domain (e.g., a fuel leak causing a crash) and relies on user insight.
- Future Work:
- Feedback Integration: Incorporating user feedback (accepted, rejected, ignored suggestions) to refine likelihood calculations for future anomalies.
- Severity and Source Data: Providing more detailed anomaly information (severity, source) to improve diagnosis accuracy.
- Automated Repair: The ultimate goal is to enable the system to automatically repair models, not just suggest repairs.
Summary updated