Curing Anomalous Extensions

Paul MorrisAAAI 1987

Abstract: In the late 1980s, the AI community faced a significant challenge known as the "Hanks-McDermott anomaly," where formal nonmonotonic logics produced counter-intuitive conclusions in temporal reasoning tasks. Paul Morris's 1987 paper, "Curing Anomalous Extensions," identifies the root cause of this issue as a lack of well-foundedness in logical justifications. By emulating the behavior of Truth Maintenance Systems (TMS) through nonnormal default rules, Morris provides a formal mechanism to eliminate these anomalies, bridging the gap between practical AI implementations and formal logical theory.

The Yale Shooting Problem and the Anomaly

In the mid-80s, the "Yale Shooting Problem" (introduced by Hanks and McDermott) exposed a fundamental flaw in how AI systems reasoned about change. In this thought experiment, a gun is loaded, a period of time passes (waiting), and then the gun is fired at a victim. Intuitively, the victim should be dead. However, standard nonmonotonic logics—designed to revise conclusions when new facts appear—produced two equally valid "extensions" or versions of reality:

  1. The Intuitive Extension: The gun remains loaded during the wait, and the victim dies.
  2. The Anomalous Extension: The gun "mysteriously" becomes unloaded during the waiting period, allowing the victim to survive.

Because the logic merely sought consistency, it treated the nonsensical unloading of the gun as a valid way to resolve the state of the world. Initially, researchers believed this was a specific failure of temporal reasoning. Morris, however, argues that the problem is more fundamental.

The Structural Root: Well-Foundedness

Morris demonstrates through comparative analysis that the anomaly is not unique to time or causality. By stripping away the temporal narrative, he shows the same issue occurs in simple cases, such as reasoning about whether a bird can fly if it might be an "abnormal" animal.

The core issue is a lack of well-foundedness. In practical Truth Maintenance Systems (TMS), every belief must have a clear, non-circular chain of justification back to a premise. In formal default logic, the mere absence of a reason not to believe something (like the gun unloading) was sometimes treated as a sufficient reason to believe it. This lack of required justification allows anomalous extensions to flourish.

Emulating TMS with Nonnormal Defaults

To "cure" the anomaly, Morris proposes a unification of Reiter’s Default Logic and Doyle’s Truth Maintenance Systems. He identifies that TMS systems avoid these traps because their justifications are unidirectional and require positive support.

He suggests representing TMS justifications as "nonnormal" default rules. Unlike "normal" defaults, which assume a proposition is true if it is simply consistent with current knowledge, nonnormal defaults allow for specific "unless" conditions that mimic the "OUT" justifiers in a TMS.

A representation of a TMS justification using default logic notation.

By reformulating the problem using these rules, Morris proves that the anomalous extension disappears. The logic is forced to follow the chain of justification, leaving only the intuitive result.

Addressing "Wishful Thinking" and Asymmetry

One of the paper's key insights is the distinction between symmetry in logic and asymmetry in reasoning. Logical implications are often symmetric (if A implies B, then Not-B implies Not-A), but human and TMS justifications are unidirectional. Morris argues that AI must embrace this asymmetry to avoid the "Wishful Thinking" problem.

For example, if an AI's goal (e.g., "the victim survives") conflicts with its defaults, a naive system might "fix" the contradiction by assuming the gun was never loaded. Morris posits that contradiction handling must be an extra-logical process that distinguishes between choices (actions), hypotheses (defaults), and observations (hard facts).

Conclusion and Future Work

Morris’s work provided a roadmap for giving TMS a solid semantic basis in default logic. While he successfully "cured" the anomaly, he noted that nonnormal rules do not always adjust smoothly to new information compared to the incremental updates of a physical TMS. He concluded that a truly robust system requires a sophisticated backtracking mechanism capable of understanding the source of contradictions, suggesting that the future of AI reasoning lies in the intersection of formal logic and strategic belief revision.

Summary updated

BibTeX
@inproceedings{morris1987curing,
  author = {Paul Morris},
  title = {{Curing Anomalous Extensions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure