Probabilistic Abduction in a Fuzzy Logic Framework

Abstract: We study the problem of explaining observations about the probabilities of events, such as "it rains $20\%$ of the time", "rain and snow are equally likely", etc. We explain these statements with a probability distribution or a statement about probabilities of (other) events that are consistent with our knowledge and entail the observation. We formalise this problem in a fuzzy probabilistic logic $\mathsf{FP}$. We define and motivate the notions of abduction problems and their solutions. Our main technical contribution is a comprehensive study of the complexity of solution recognition and existence for a given abduction problem in $\mathsf{FP}$ for the case of full language and its disjunctive-clause fragments. We also obtain a translation of classical probabilistic abduction (finding the most likely explanation of a given event) to $\mathsf{FP}$.

• The paper introduces a fuzzy probabilistic logic framework that extends Łukasiewicz logic to include probabilistic abduction and real-valued truth assignments.
• It provides detailed complexity characterizations, showing NP-, DP-, and Σ₂^p-completeness for various solution recognition and existence tasks.
• The study translates classical abduction into the FP framework, enabling principled explanation selection with maximal entropy and minimality criteria.

Summary of Probabilistic Abduction in Fuzzy Logic

Motivation and Context

Probabilistic abduction seeks the explanation for observations about event frequencies or probabilities, such as statements relating to weather occurrences ("it rains 20% of the time", "rain and snow are equally likely"). Existing frameworks in probabilistic logic programming, Bayesian and Markov logic networks represent a range of abductive reasoning approaches, but they often lack a rigorous logical formalization grounded in uncertainty and fuzzy logics. This paper addresses probabilistic explanation under uncertainty, formalizing abduction in fuzzy probabilistic logic (FP), an expansion of Łukasiewicz logic for real-valued degrees of truth.

Formal Framework

FP extends Łukasiewicz logic by introducing probabilistic atoms, i.e., expressions $Pr(\phi)$ where $\phi$ is a propositional formula and $Pr(\phi)$ denotes the probability of $\phi$. Compound probability expressions are linked via Łukasiewicz connectives ($\neg$, $\odot$, $\oplus$, $\rightarrow$, etc.). Modal extensions include the Baaz $\triangle$ operator and truth constants. The semantics are defined via probability distributions over subsets of variables, aligned with classical event sets but allowing real-valued truth assignments. The FP framework generalizes classical entailment and satisfiability notions to real-valued logic, leveraging NP-completeness for satisfiability and coNP-completeness for entailment (2604.22064).

Definition and Classification of Abduction in FP

An abduction problem in FP (FP AP) is a tuple $P = \langle \Gamma, \delta, H, C \rangle$, where $\phi$0 is the background probabilistic theory, $\phi$1 is the observation, $\phi$2 is a finite set of hypotheses (terms), and $\phi$3 is a finite set of values (e.g., rational steps between 0 and 1). Solutions are defined in two classes:

• Sufficient Solution: A probabilistic interval term expressing values for the hypotheses that, together with the theory, entail the observation.
• Full Solution: An assignment specifying a probability distribution over all events/hypotheses, compatible with the theory and observation, and analogous to a maximally specific explanation.

The framework accommodates non-truth-functional probabilities, i.e., the probability of $\phi$4 cannot generally be derived from $\phi$5 and $\phi$6 due to dependencies in complex events.

Complexity Results

The paper rigorously studies three main computational tasks:

• Solution Recognition: Checking if a candidate explanation solves an abduction problem.
• Solution Existence: Determining whether any solution exists.
• Minimality and Maximal Entropy Criteria: Preferring explanations based on logical minimality or probability distribution entropy.

The main complexity results are:

Task	Complexity	Notes
Sufficient Solution Recognition	DP-complete	Same as classical/Łukasiewicz logic
Full Solution Recognition	Polynomial	Unique distribution simplifies checking
Sufficient Solution Existence	$\phi$7-complete	Harder due to uncertainty in probability assignments
Full Solution Existence	NP-complete	Only one probabilistic model needs to be checked
Concise Entropy-Maximal Solution	coNP-complete	Highest entropy among concise full solutions
Entailment-Minimal Solution	DP-complete	Logical minimality criterion

A systematic translation is given from classical probabilistic abduction (as in [Poole1993], [Sato1995]) to FP, allowing preferred solutions to be identified as those with maximal probability (see details on conditional probability characterization and the formal translation).

Fragment Analysis

Restricting the FP language yields tractable subsets:

• Probabilistic Simple Clauses (PSC): Fragments where theories are disjunctions of non-nested probabilistic atoms. Solution recognition and existence are NP-complete, minimality is DP-complete.
• Probabilistic Interval Clauses (PIC): Fragments supporting interval constraints on event probability. Complexity matches arbitrary FP-theories unless further syntactic “cover-free” restrictions are employed.
• SPCF (Short Probabilistic Cover-Free) Theories: Enables polynomial-time sufficient solution existence, but full solution existence remains NP-complete.

These fragments generalize tractable fragments in classical and fuzzy logics, enabling syntactic restrictions akin to (CPL) Horn clauses for polynomial abductive reasoning.

Prioritization of Explanations

Logical minimality is generalized to FP-minimality (entailment-minimal). For full solutions, the principle of maximal entropy is applied, favoring explanations that “spread” probability assignments as uniformly as possible within the constraints. This matches information-theoretic principles and is shown to be coNP-complete to recognize among concise solutions.

Translation of Classical Probabilistic Abduction

Classical abduction problems with preference assignments or probability distributions are translated to FP-theories expressing the assigned probabilities as constraints. Verification of solution existence or recognition corresponds to checking the coherence of probability assignments and entailment of the observation, with DP-completeness and, when assignments fully specify a distribution, coNP-completeness.

Implications and Future Directions

• Practical Impact: The framework formalizes probabilistic abduction in settings where observations are frequencies or probability statements, such as diagnosis, scientific reasoning, and explainable AI. It subsumes classical probabilistic abduction and logic programming variants.
• Theoretically: It rigorously connects fuzzy logic, probabilistic modal logic, and abduction, mapping the boundary between tractable and intractable reasoning for sublanguage fragments.
• Contradictory Claim: It is strictly harder to check existence of sufficient solutions than full solutions under reasonable encoding assumptions.
• Numerical Results: Complexity bounds are precisely established for recognition and existence scenarios, e.g., DP-completeness for sufficient solution recognition and $\phi$8-completeness for existence.

The paper suggests extensions to abduction with conditional probabilities, more expressive fuzzy logics, relevance and necessity checking, translation to fuzzy propositional abduction, and prioritization criteria alternative to entropy (e.g., maximum likelihood). Future work may further investigate contrastive explanations, extended abduction, or hybrid uncertainty models (e.g., belief functions, qualitative probabilities).

Conclusion

This research delivers a comprehensive logical and computational foundation for probabilistic abduction in fuzzy logic settings. It provides precise complexity characterizations, formal translations between frameworks, and prioritization mechanisms for explanations. The theoretical findings enable principled, tractable reasoning about probabilistic observations and explanations, expanding the capabilities of AI systems for diagnosis, explanation, and uncertainty reasoning. The possible extensions to conditional probability, qualitative logics, and new minimality criteria promise further integration with advanced reasoning and learning paradigms (2604.22064).