TL;DR
This paper introduces Horizon, a novel decision aid that uses evidential reasoning to fuse uncertain and incomplete information for naval target classification. It addresses the challenge of information overload faced by sonar operators, providing a more accurate and timely identification of potential threats.
What Problem Does This Paper Address?
Naval sonar operators are tasked with a critical mission: detecting, localizing, and classifying potential targets. This process relies on a deluge of data from various sensors, which is often uncertain, incomplete, and even contradictory. As sensor technology advances, the sheer volume of this information risks overwhelming the operator, leading to potential misclassifications or delayed responses. The paper tackles this "information overload" problem by developing a decision aid that can intelligently fuse disparate pieces of evidence to provide a reliable target classification.
Methods
The core of Horizon is evidential reasoning (E-R), a formalism that extends classical probability theory. E-R is particularly adept at handling uncertain, incomplete, and error-prone information by maintaining associations between belief measures and disjunctions of events, rather than forcing probabilities onto atomic possibilities.
Horizon builds upon existing E-R techniques by:
- Representing and Fusing Imprecise Information: It uses "frames of discernment" to define different attributes of a target (e.g., type, nationality, speed) and "basic probability assignments" (BPAs) to express confidence in propositional statements within these frames.
- Novel Fusion Algorithms: While employing Dempster's rule of combination, Horizon also introduces a new algorithm for combining dependent evidence, a crucial aspect in real-world scenarios where sources might not be entirely independent. It also incorporates Smets' algorithm, which offers a conservative weighting of evidence suitable for evolving military domains.
- Automated Discounting: Horizon includes an "auto-discounting" feature that automatically adjusts the confidence of incoming evidence based on the perceived credibility of its source. This helps to minimize human biases and account for the reliability of sensors or intelligence reports.
- Graphical User Interface (CR-Editor): A user-friendly interface allows for the graphical representation and editing of "compatibility relations" between frames, making the knowledge engineering process more accessible.
Datasets
Not discussed in the paper. The paper mentions using synthetic data for evaluation and drawing information from sources like "Jane's Fighting Ships" for domain knowledge.
Key Insights and Ideas
- Evidential Reasoning is Well-Suited for Naval Classification: The paper strongly advocates for E-R as a robust methodology for fusing high-level, uncertain information in tactical environments.
- The Challenge of Evidence Independence: While E-R typically assumes independence, the paper acknowledges this is a controversial area and a significant practical challenge, especially with intelligence reports. Horizon's inclusion of a dependent evidence fusion algorithm is a key innovation to address this.
- Balancing Automation and Operator Control: Horizon aims to be a decision aid, not a replacement for the operator. The auto-discounting feature, while useful, can reduce information content, necessitating operator training to know when to override it.
- Transparency is Key for Trust: The "point-and-click" interface and the planned explanation facility are crucial for building operator confidence in the system's recommendations.
Key Results
- Demonstrated Suitability of E-R: Preliminary results suggest that evidential reasoning is an appropriate technique for high-level information fusion in target classification.
- Performance on Synthetic Data: For a worst-case scenario with 35 BOEs and an average of 211 propositions, Horizon took 2.3 minutes to perform 25 discounting, 29 translation, and 35 fusion operations on a Pentium 133. This is considered reasonable for some tactical situations.
- Qualitative Findings: The system's graphical interface is commended for its ease of use and provision for rapid interpretation of information.
Limitations and Future Work
- Optimisation for Speed: Horizon is currently a technology demonstrator and is not optimized for speed or efficiency.
- Dependent Evidence Fusion: The utility of the dependent evidence fusion algorithm is still under investigation.
- Operator Training: While the interface is user-friendly, some level of training is still required for operators to effectively distribute evidence and ensure its independence.
- Auto-Discounting Nuances: The paper notes that auto-discounting can lead to inconclusive results in certain situations, suggesting operators should be trained to redo fusion without it when necessary.
- Further Investigation into Confidence Values: The appropriateness of the chosen discount rates for "certain," "probable," and "possible" confidence levels is still being investigated.
Summary updated