Extended Plausible Inference

Leonard Friedman

Introduction

Leonard Friedman’s "Extended Plausible Inference" (PI) represents a significant advancement in artificial intelligence, bridging the gap between rigid propositional logic and the heuristic-based uncertainty models of early expert systems. By integrating the mathematical foundations of Dempster-Shafer theory, the PI system enables sophisticated diagnostic reasoning for complex, real-world systems where certainty is a luxury.

The Limitations of Early Expert Systems

Standard logic operates on a binary of certainty: a predicate is either true or false. In fields like medicine or aerospace engineering, however, evidence is often incomplete or conflicting. Early expert systems like MYCIN and PROSPECTOR attempted to model "degrees of belief," but their underlying formulas were frequently "ad hoc." These systems struggled with conflicting evidence and failed to account for set inclusion—where one assertion is a subset of another.

Friedman identified the need for a generalized theory of plausible inference that could handle arbitrary propositional calculus expressions while remaining mathematically sound. The goal was a system capable of not only confirming hypotheses but also denying them, handling contradictions, and intelligently focusing its "attention" within massive search trees.

The PI System: Mathematical Foundations

The PI system extends traditional logic through several key innovations:

  • Four Modes of Reasoning: Unlike the two modes of standard logic, PI utilizes four: modus ponens (confirmation), modus tollens (denial), and their counterparts for propagating disbelief.
  • Dempster-Shafer Integration: PI incorporates rigorous mathematical treatments of evidence to correct summation errors found in earlier systems. This allows for a clear distinction between consonant evidence (subsets) and heterogeneous evidence (potentially conflicting).
  • Relevance Factors ($Al$): Links between antecedents and consequents are weighted by relevance factors. PI includes a heuristic to estimate these factors from evidence, allowing the system to learn from experience rather than relying solely on static expert estimates.
  • Belief vs. Disbelief: PI treats belief ($I$) and disbelief ($D$) as independent quantities. This better models human reasoning, where a lack of evidence for truth does not automatically imply evidence for falsity.

Bi-directional Propagation and Meta-Reasoning

A defining feature of the PI system is its ability to propagate belief bi-directionally. While earlier systems typically reasoned from evidence toward a conclusion, PI allows changes in certainty to flow in both directions through a semantic network.

Comparison of belief propagation in MYCIN/PROSPECTOR vs. the PI system.

For large-scale problems, such as a spacecraft with thousands of components, exhaustive search is computationally prohibitive. PI employs "meta-rules"—rules about rules—to select specific "neighborhoods" of the logic tree to test. This allows the system to prioritize tests based on criteria like cost or expected information gain, effectively pruning irrelevant branches of the search space.

Application: Spacecraft Fault Diagnosis

The PI system was evaluated using a simplified problem drawn from Jet Propulsion Laboratory (JPL) spacecraft operations. The knowledge base was derived from models of "normal operation" generated by automated planners.

Logic network for a spacecraft signal reception diagnostic task.

In this domain, the system demonstrated the ability to isolate a "failed leaf" in a plan tree by pruning branches confirmed to be functioning correctly. This narrowed down the potential causes of failure efficiently, even in the presence of complex causal relations.

Conclusion

Extended Plausible Inference provides a mathematically rigorous framework for reasoning under uncertainty. By moving beyond binary logic and one-directional propagation, it offers a more flexible and powerful tool for complex diagnostic tasks. While initial implementations were limited in scale, the principles of meta-reasoning and bi-directional belief propagation remain foundational to modern AI diagnostic systems.

Summary updated

BibTeX
@inproceedings{friedman1981extended,
  author = {Leonard Friedman},
  title = {{Extended Plausible Inference}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}

Loading related content...

Figure