A New Framework for Sensor Interpretation: Planning to Resolve Sources of Uncertainty

Norman Carver, Victor R. LesserAAAI 1991

Abstract

Sensor interpretation in complex environments is a constructive problem characterized by massive search spaces and inherent data ambiguity. Traditional blackboard systems often fail to efficiently resolve competing hypotheses, relying on passive evidence accumulation. This paper introduces RESUN (REsolving Sources of UNcertainty), a framework that transforms sensor interpretation into an active, strategic planning process. By representing uncertainty symbolically through "Sources of Uncertainty" (SOUs), RESUN enables sophisticated reasoning strategies such as differential diagnosis. Experimental results in a simulated aircraft monitoring domain demonstrate that goal-directed control and uncertainty resolution can reduce CPU time by up to 54% while achieving higher interpretation confidence.

Introduction

The challenge of sensor interpretation lies in transforming low-level, noisy data—such as radar pings or acoustic signals—into high-level, coherent narratives like vehicle tracks or mission profiles. This is a constructive task where the number of possible interpretations is effectively infinite. Traditional systems typically employ a "hypothesize and test" loop, which struggles with ambiguity; when multiple hypotheses explain the same data, these systems often continue gathering evidence for all alternatives without a clear strategy for pruning the search space. To achieve efficiency in real-time monitoring, a system must understand the qualitative nature of its uncertainty to actively resolve it, rather than simply waiting for more data to accumulate.

Methods

The RESUN framework is built upon three primary pillars designed to move beyond numeric probability scores:

  1. Symbolic Evidential Representation: Instead of assigning a single black-box probability (e.g., 0.75), RESUN utilizes symbolic labels called Sources of Uncertainty (SOUs). These labels, such as partial-evidence, possible-alternative-explanation, or negative-evidence, allow the system to distinguish between a lack of data and competition between theories.
  2. Incremental Script-Based Planner: Control logic is implemented as a hierarchical planner that treats uncertainty resolution as a top-level goal. This allows the system to select the most appropriate interpretation method dynamically based on the specific SOUs present.
  3. Refocusing Mechanism: This allows the planner to postpone difficult decisions until sufficient evidence is gathered for a meaningful comparison, or to opportunistically switch focus when critical new data arrives.
Symbolic representation of track extensions and associated Sources of Uncertainty (SOUs).

A key strategy enabled by this framework is differential diagnosis. If two hypotheses explain the same data, RESUN can actively seek evidence that specifically contradicts one, thereby indirectly strengthening the other.

Results

The framework was evaluated using a simulated aircraft monitoring application involving complex sensor artifacts like ghosting, noise, and sensor errors. Four experiments were conducted to measure the evolution of system efficiency as more sophisticated control strategies were introduced.

Performance metrics across four experimental configurations showing reductions in CPU time and action counts.

Key findings include:

  • Efficiency Gains: Moving from basic strategies to goal-directed control (Experiment 2) reduced CPU time by 26%, with reductions reaching up to 54% in more complex scenarios.
  • The Power of Disproof: Implementing differential diagnosis (Experiment 4) further reduced CPU time by 24-28% while simultaneously reaching higher confidence levels in final interpretations.
  • Strategic Flexibility: The system successfully adapted its behavior based on goals, such as increasing effort to disprove alternatives when higher certainty thresholds were required.

Conclusion

RESUN demonstrates that treating uncertainty as a qualitative, resolvable entity significantly improves the efficiency and effectiveness of sensor interpretation. By labeling the nature of doubt, the system can apply targeted reasoning tools to prune search spaces. While the planning overhead is a consideration for real-time systems, the strategic flexibility and reduction in unnecessary inference actions provide a substantial net benefit. Future work includes optimizing the implementation and extending the framework to diverse domains such as domestic acoustic monitoring.

Summary updated

BibTeX
@inproceedings{carver1991new,
  author = {Norman Carver and Victor R. Lesser},
  title = {{A New Framework for Sensor Interpretation: Planning to Resolve Sources of Uncertainty}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure