Introduction
In the early 1980s, the field of Artificial Intelligence was deeply rooted in formal logic. Researchers sought to model human intelligence by refining deductive systems, often attempting to "fix" the inherent messiness of real-world reasoning by introducing probabilistic truth values or non-monotonicity. In "Simulating Non-Deductive Reasoning," James W. Hearne argues that this approach is fundamentally flawed. Rather than viewing non-deductive reasoning—the everyday guesswork and decision-making humans excel at—as a "leaky" or broken version of a mathematical proof, Hearne proposes that it is a distinct process governed by explanatory coherence and situational values.
The Problem of Detachment
The central challenge identified by Hearne is the Problem of Detachment. In a purely deductive system, a conclusion is reached when the premises necessitate it. However, in the real world, an intelligent agent is faced with an infinite array of possible explanations and rival hypotheses. The logical "loop" of gathering evidence could, in theory, continue forever.
The critical question for AI is: how does a system decide when to stop calculating and actually "commit" to a belief or an action? This "detachment" from the reasoning process is not a logical necessity but a strategic requirement.
The Method of Procedure
Hearne moves away from rigid algorithms toward a "Method of Procedure" for evidential reasoning. This framework consists of a six-step iterative strategy designed to manage competing hypotheses:
- Identify: Recognize a fact or phenomenon that requires an explanation.
- Articulate: Generate potential hypotheses that could account for the fact.
- Rank: Evaluate these hypotheses based on their initial plausibility.
- Search: Actively seek out new data that might confirm or refute the leading candidates.
- Iterate: Return to the ranking or searching phases as new information arrives.
- Detach: Select the most coherent conclusion and proceed to act upon it.
Unlike standard deductive algorithms, this process lacks a built-in "halt" command. The decision to move to step six is driven by external constraints—time, resources, and the consequences of the decision—rather than the internal logic of the data alone.
Value-Dependent Detachment
Perhaps the most significant contribution of Hearne’s work is the concept of Value-Dependent Detachment. To illustrate this, he draws on an example from philosopher Michael Scriven. Consider a person hauling a bale of hay into a loft; they might "detach" the conclusion that the rope is strong enough after a single, quick tug. However, if that same person were hauling their daughter into the loft, the threshold for evidence would be significantly higher.
The underlying logic of the rope's tensile strength remains the same in both scenarios. What changes is the value associated with the outcome. Hearne argues that a truly intelligent system must be able to adjust its reasoning thresholds based on the "consequences of being wrong." Detachment is thus a rational response to situational demand.
Explanatory Coherence and Efficiency
Hearne posits that "explanation" is the glue of intelligence. We do not judge a conclusion in isolation; we judge it by how well it "coheres" with our existing knowledge base. This leads to a redefinition of "comprehension." Citing Scriven’s "comprehension theorem," Hearne links the abstract concept of understanding to computational efficiency. In an environment where an agent cannot possibly remember every discrete fact, comprehension allows for the efficient storage and retrieval of information by organizing it into coherent, explanatory structures.
Conclusion
"Simulating Non-Deductive Reasoning" provides a qualitative framework that bridges the gap between philosophical inquiry and computer science. By identifying plausibility judgments as products of explanatory coherence and defining detachment as a value-based strategy, Hearne challenges the AI community to look beyond formal logic. While the paper does not provide a specific code implementation, it sets the stage for a generation of AI that prioritizes strategic decision-making over exhaustive deduction.
Summary updated