Introduction
In the late 1980s, artificial intelligence researchers grappled with the challenge of non-monotonic reasoning—the human ability to "jump to conclusions" and retract them when new evidence arises. While frameworks like McCarthy’s Circumscription and Reiter’s Default Logic provided initial tools, they often lacked a rigorous semantic foundation and failed to resolve classic paradoxes. Hector Geffner’s "On the Logic of Defaults" introduced a system, logic $L$, that grounds default reasoning in probability theory and a formal notion of "relevance."
The Framework of Logic $L$
Geffner’s system distinguishes between two types of knowledge:
- Background Context ($K$): Universal truths and general defaults (e.g., "Birds fly").
- Evidential Set ($E$): Specific facts about the current scenario (e.g., "Tweety is a penguin").
The logic is governed by six inference rules. Rules 1-4 are derived from Ernest Adams’ logic of conditionals, handling basic logical properties like disjunction. Rules 5 and 6 introduce the Irrelevance Predicate $I(\cdot)$, which allows the system to maintain a default conclusion as long as new evidence does not provide a specific "argument" for its failure.
Specificity and Conflict Resolution
One of the primary strengths of logic $L$ is its ability to handle conflicting defaults through specificity. In earlier systems, manual priorities were often required to ensure that specific information (e.g., "Penguins don't fly") overrode general information (e.g., "Birds fly"). In $L$, this behavior emerges naturally from the mathematical structure of the rules.
As shown in Figure 1, the system can represent complex relationships where different paths lead to conflicting conclusions. By treating defaults as constraints true with a probability "infinitesimally close to one," $L$ provides a principled way to resolve these conflicts without arbitrary weighting.
Solving the Yale Shooting Problem (YSP)
The Yale Shooting Problem was a notorious benchmark where temporal reasoning systems failed to conclude that a victim would die after being shot with a loaded gun. Earlier logics often allowed "weird" interpretations where the gun spontaneously became unloaded during a waiting period to preserve the victim's "alive" status.
Geffner solved this by introducing the Causal Fork. By interpreting relevance as graph separation (similar to Bayesian networks), the logic prevents information from "flowing backwards" through causal actions.
In Figure 2, the sequence of states $A(0)$ through $A(3)$ represents the victim's status. The introduction of the shooting action $S(2)$ with a loaded gun $L(2)$ provides a definitive argument for the transition to a non-alive state, correctly overriding the default persistence of life.
Conclusion and Legacy
Logic $L$ successfully bridged the gap between symbolic AI and probabilistic reasoning. While it faced challenges regarding computational complexity and the rigid boundary between background and evidential knowledge, it laid the groundwork for modern causal reasoning and Bayesian approaches to commonsense. By formalizing "relevance," Geffner provided a roadmap for how machines might one day navigate the messy, uncertain world of human logic.
Summary updated