Towards Situated Explanation

Raja Sooriamurthi, David B. LeakeAAAI 1994

Introduction

Traditional artificial intelligence explanation systems have historically operated in a logical vacuum. When a system fails or encounters an anomaly, it typically generates a post-mortem analysis—a static logical trace of the failure—regardless of the immediate needs of the agent or the constraints of the environment. This paper argues that for AI to be truly effective in dynamic settings, explanation must be situated.

Situated explanation is defined as a goal-driven, utility-based process that adapts to environmental constraints and the system's immediate operational needs. Rather than seeking logical completeness as an end in itself, a situated explainer treats the act of explaining as a functional tool to help the system resume its tasks or pivot to new ones.

The Problem: Context-Free Explanations

In early AI research, an explanation was often treated as a neutral output. However, the authors posit that the "best" explanation is entirely dependent on the context. Consider a robot whose car fails to start while it is attempting to catch a flight. The utility of a specific explanation changes based on several real-world factors:

  1. Time Constraints: If a flight departs in 20 minutes, a deep mechanical diagnosis of the fuel injection system is useless. The system simply needs to know the car is unreliable so it can immediately call a taxi.
  2. Information Gaps: Systems rarely possess perfect information. They must decide which pieces of data are worth the "cost" (in time or energy) to acquire.
  3. Shifting Goals: The depth of an explanation depends on whether the goal is immediate recovery (getting to the airport), anomaly resolution (understanding the failure), or long-term prevention (fixing the car for future use).

Proposed Methodology

The authors propose an integrated architecture that combines a planner (operating in a simulated environment) with a case-based explainer. This model relies on four key technical pillars:

Case-Based Reasoning (CBR)

Instead of generating explanations from first principles—which is computationally expensive—the system retrieves and adapts prior explanations from similar past problems. This provides a "cheap" initial hypothesis that can be refined if necessary.

Knowledge Planning

The search for information is treated as a planning problem. If the system hypothesizes that a battery is dead, it generates a plan to "check the battery." This allows the system to weigh the action of gathering information against other physical actions.

Utility-Based Reasoning

The system assigns a "utility" value to different explanatory paths. It evaluates whether the effort required to reach a specific level of explanatory detail is worth the potential benefit to the current high-level plan. If the cost of diagnosis exceeds the benefit of the information, the search is truncated.

Hierarchical Explanation

Explanations are produced at varying levels of abstraction. Depending on the system's needs, it may provide a vague summary ("engine trouble") or a highly specific diagnosis ("corroded battery terminal").

Scenario-Based Implementation

The model was tested using a scenario-based implementation in a simulated world. The primary test case involves a planner attempting to reach an airport. When the car fails to start, the system must choose between diagnosing the vehicle or pivoting to an alternative plan, such as taking a taxi.

This simulation allows for the testing of opportunistic adaptation. For instance, if the system learns that the flight has been delayed, it might "opportunistically" use the newly available time to perform a deeper mechanical diagnosis that it had previously skipped due to time pressure.

Key Insights and Results

The research yields several qualitative insights into the nature of situated reasoning:

  • Explanation as a Means to an End: The ultimate metric for an explanation is the "successful functioning of the system," not just logical accuracy.
  • Packaging Knowledge Actions: CBR is used not just to find answers, but to "package" the steps needed to find those answers. Prior cases guide the system on what information was useful in the past, streamlining the current information search.
  • Contextual Depth: Context determines the depth of the search. This prevents the system from becoming bogged down in irrelevant details when time is of the essence, ensuring that the explanation process remains aligned with the agent's primary goals.

Limitations and Future Work

As a model under development, the current implementation is limited to a simple simulated world. Its performance in high-dimensional, noisy, or open-ended real-world environments remains a subject for future study. Additionally, the authors identify the need to better handle "mid-thought" interruptions—how the explanation process should gracefully update when environmental conditions change while the system is in the middle of a diagnostic plan. Finally, the scalability of the case base remains an open question, as a massive library of prior cases would require more sophisticated prioritization mechanisms.

Summary updated

BibTeX
@inproceedings{sooriamurthi1994towards,
  author = {Raja Sooriamurthi and David B. Leake},
  title = {{Towards Situated Explanation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure