Diagnosis of Multiple Faults in a Nationwide Communications Network

Thomas L. Williams, Paul J. Orgren, Carl L. Smith

Abstract: This paper presents the Network Diagnostic System (NDS), an early expert system designed to troubleshoot complex nationwide communications networks (COMNET). Unlike contemporary diagnostic AI that typically assumed single-fault scenarios, NDS introduced a framework for identifying multiple, dependent, and intermittent faults. By utilizing a refinement hierarchy and a mechanism for evidence revocation after repairs, NDS successfully automated the reasoning of intermediate-level technicians, providing a scalable solution for network maintenance in the early 1980s.

Introduction

Published in 1982, this research describes the development of the Network Diagnostic System (NDS). At a time when digital and analog hardware were becoming increasingly integrated into vast "virtual wires," troubleshooting required navigating a complex web of telecommunication processors, statistical multiplexors, modems, and cables. NDS was built using the ARBY expert system shell to move beyond the limitations of "single-fault" assumptions in large-scale network troubleshooting.

The Challenge of Nationwide Networks

In the early 1980s, networks like COMNET faced three major diagnostic hurdles that traditional methods could not resolve:

  1. Dependent Faults: A failure in one component (e.g., a multiplexor reset) could trigger secondary failures elsewhere (e.g., a modem losing its baud rate).
  2. Independent Faults: Unrelated issues occurring simultaneously, such as a user damaging a terminal while attempting to fix a network-side modem problem.
  3. Intermittent Faults: Transient errors caused by telephone line noise that did not fit a simple binary "working/broken" state.

Human experts capable of navigating these complexities were scarce, necessitating an automated system that could perform at least at an "intermediate" level to isolate Field Repairable Units (FRUs).

Methodology: The Virtual Wire and Refinement Hierarchy

The core of NDS is the "virtual wire" abstraction, which treats geographically distributed paths as linear sequences of electronic devices. This simplification allowed the system to manage a massive networking problem as a manageable diagnostic chain.

Network topology illustrating the "virtual wire" abstraction across different geographic locations from Houston to Brooklyn.

NDS treats diagnosis as a heuristic search through a "hypothesis space." It begins with a broad hypothesis (e.g., "The entire communication path is down") and uses diagnostic tests, primarily loop-back tests, to narrow the search to specific components.

Evidence Revocation and Post-Repair Logic

The paper's most significant logical contribution is its mechanism for "evidence revocation." When a component is repaired, the system does not simply terminate. Instead, it revokes all evidence associated with that component and runs a termination test. If the network remains down, NDS resumes its search from the most specific hypothesis that was not invalidated by the repair. This allows the system to systematically peel back layers of multiple faults.

The Hypothesis Refinement Procedure flowchart, detailing the iterative diagnostic and repair loop used by NDS.

Key Results and Insights

NDS demonstrated that expert systems could achieve human-level performance in niche technical domains with relatively low overhead. Key outcomes included:

  • Performance: The system performed at the level of an intermediate COMNET diagnostician.
  • Efficiency: Development required only four person-months of knowledge engineering and one person-month of a domain expert's time.
  • Rule Integration: The system successfully managed 150 ARBY rules and 12 different diagnostic tests to navigate its refinement graph.

Limitations and Future Work

Despite its success, the authors identified the "knowledge engineering bottleneck" as a primary hurdle, suggesting that future systems should aim to automatically generate refinement hierarchies. Additionally, the system's reliance on high-end environments at the time highlighted a need for portability to the emerging generation of microcomputers. The project underscored a classic AI "catch-22": the domains most in need of AI are often those where human experts are too busy to assist in building the system.

Summary updated

BibTeX
@inproceedings{williams1983diagnosis,
  author = {Thomas L. Williams and Paul J. Orgren and Carl L. Smith},
  title = {{Diagnosis of Multiple Faults in a Nationwide Communications Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure