(1)
Journal of Fuzzy Systems and Control, Vol. 4, No 1, 2026 |
Causal-Entropic Fuzzy Inference: A Bayesian Framework for Explainable and Robust Reasoning
Jin-Hyok Choe 1,*, Yon-Ju Jang 2, Son-Il Kwak 3, Ok-Sim Ri 4, Hyon-U Kong 5
1, 2, 3, 5 Department of Artificial Intelligence, Kim Il Sung University, Pyongyang, Democratic People’s Republic of Korea
4 Department of Foreign Languages and Literature, Kim Il Sung University, Pyongyang, Democratic People’s Republic of Korea
Email: 1 cjh920321@163.com, 2 yj.jang@ryongnamsan.edu.kp, 3 si.kwak@ryongnamsan.edu.kp, 4 os.ri@ryongnamsan.edu.kp,
5 khu1210@163.com
*Corresponding Author
Abstract—Traditional fuzzy reasoning methods exhibit limitations in satisfying the reductive property and handling uncertain environments. This paper proposes a novel Causal-Entropic Fuzzy Inference (CEFI) framework that integrates causal discovery with Bayesian inference to overcome these limitations. The proposed method consists of three main components: (1) a causal rule discovery mechanism based on conditional independence tests, (2) an entropic inference engine utilizing variational free energy minimization, and (3) an active perception module for strategic information gathering. Experimental results on SISO and MISO systems demonstrate that CEFI achieves 99.4% reductive property, outperforming state-of-the-art methods by 7.3-31.2% in noisy environments while providing causal explanations for reasoning processes.
Keywords—Fuzzy Reasoning; Causal Discovery; Bayesian Inference; Free Energy Principle; Explainable AI; Reductive Property
The field of fuzzy reasoning has witnessed substantial evolution since Zadeh's pioneering work on fuzzy sets and the compositional rule of inference (CRI) [1], [2]. Early research focused primarily on extending traditional logical frameworks to handle uncertainty and imprecision, with Zadeh's CRI establishing the fundamental mechanism for approximate reasoning using min-max operations. While this approach demonstrated success in various control applications [3], it soon became apparent that CRI suffered from limitations in satisfying the reductive property, particularly in multi-rule systems [4].
The pursuit of improved reductive properties motivated the development of similarity-based reasoning methods. Turksen and Zhong [5] proposed approximate analogical reasoning using similarity measures and interval-valued fuzzy sets, while Wang et al. [6] developed fuzzy similarity inference methods that showed improved performance in specific domains. These approaches focused on measuring the degree of resemblance between premises and rule antecedents, though they often encountered challenges with information loss during similarity calculations and inconsistent performance across different reasoning contexts.
Implication-based methods emerged as another significant research direction. Wang [7] introduced the triple implication principle (TIP) with total inference rules, followed by Zhou et al. [8] advancing the quintuple implication principle (QIP). These methods employed sophisticated implication operators to strengthen the logical foundations of fuzzy reasoning, demonstrating improved theoretical properties while maintaining graded truth values. However, practical implementations sometimes produced results that, despite mathematical soundness, failed to align with human intuition in complex reasoning scenarios.
The paradigm shifted significantly with the emergence of distance-based optimization methods. Kwak et al. [9] developed the KSI-based method utilizing moving differences and iterative optimization, while simultaneously introducing the distance measure method (DMM) employing Euclidean distance calculations with sign vectors [10]. These approaches reformulated fuzzy reasoning as an optimization problem, achieving substantial improvements in reductive property scores and establishing new performance benchmarks. The optimization-based framework represented a fundamental departure from earlier logical and similarity-based approaches, focusing on minimizing distances between premises and antecedents while satisfying specific constraints.
Despite these advances, research by various scholars [11], [12] has identified persistent limitations in distance-based methods. Their geometric foundations create inherent sensitivity to noise and measurement uncertainty, while computational complexity limits scalability in systems with large rule bases. Most critically, these methods provide limited explanatory capability, offering little intuitive insight into their reasoning processes a significant drawback in applications requiring transparent decision-making.
Recent research has explored integrating fuzzy logic with other mathematical frameworks to address these limitations. Pearl's causal inference framework [13] and Bayesian network approaches [14] have shown promise in handling uncertainty through probabilistic reasoning. The free energy principle from theoretical neuroscience [15] has inspired new variational methods for robust inference. Meanwhile, research in cognitive science [16] has highlighted the importance of causal reasoning in human cognition, suggesting new directions for fuzzy system development.
The comprehensive review of existing literature reveals substantial research gaps in the current landscape of computational reasoning systems. There exists a fundamental disconnect between the geometric foundations of current fuzzy reasoning methods and the causal-probabilistic nature of human reasoning. Current integration attempts between different reasoning paradigms typically treat them as separate computational layers with limited interaction.
This paper addresses these gaps with the following contributions:
The remainder of this paper is organized as follows. Section 2 reviews related work and establishes theoretical foundations. Section 3 presents the CEFI framework with complete mathematical details. Section 4 describes the experimental setup and presents results with analysis. Section 5 concludes the paper with a discussion of limitations and future work.
The field of fuzzy reasoning began with Zadeh's groundbreaking work on fuzzy set theory in 1965 [1], which established the mathematical basis for handling graded membership and linguistic imprecision. The subsequent development of the Compositional Rule of Inference (CRI) by Zadeh [2] provided the first systematic approach to approximate reasoning.
The fundamental CRI formulation for Fuzzy Modus Ponens (FMP) established the core mathematical relationship that enables approximate reasoning with fuzzy sets. This formulation expresses the inference process as [2], [3]:
| (1) |
In this comprehensive formulation, the operator ⊗ represents a t-norm that systematically combines the premise membership with the implication relationship, while → denotes a fuzzy implication operator that mathematically captures the logical relationship between antecedent and consequent [4]. The choice of specific t-norm and implication operators significantly influences the system's reasoning characteristics, with different combinations exhibiting varying properties in terms of inference behavior, computational efficiency, and alignment with human reasoning patterns.
For practical computational implementations using discrete fuzzy sets, this general formulation transforms into the more specific and computationally tractable max-min composition [3]:
B∗=A∗∘R | (2) |
where the matrix R encodes the fuzzy relation representing the rule "If A then B" through carefully selected implication operators [2]. This formulation enables efficient computation while maintaining the theoretical foundations of the continuous case, making it suitable for real-time applications and embedded systems where computational resources may be constrained.
The evaluation of reasoning performance in fuzzy systems has been traditionally quantified through the reductive property criteria, which systematically measure how precisely the system's conclusion matches the expected result when premises perfectly align with rule antecedents [5], [6]. This essential performance metric is formally defined through the equation [7]-[9]:
| (3) |
Early foundational investigations by Mizumoto [4], [18] provided exhaustive analyses of diverse implication operators for CRI, systematically comparing Mamdani, Gödel, Łukasiewicz, and other implication functions across varied reasoning scenarios. These meticulous studies revealed that while CRI demonstrated robust performance in simple control applications and single-rule systems, it encountered significant challenges in maintaining consistent reductive property in multi-rule systems and complex reasoning scenarios involving multiple interacting variables with non-linear relationships.
The identified limitations of basic CRI stimulated substantial research into optimization-based approaches. The emergence of distance-based optimization methods constituted a genuine paradigm shift in fuzzy reasoning research, fundamentally reconceptualizing the nature of fuzzy inference from a purely logical process to a sophisticated optimization problem. This transformative approach was pioneered by researchers through the development of methods that introduced innovative concepts of moving differences and iterative optimization to achieve unprecedented reductive property scores.
The core innovation of distance-based methods resides in their formulation of reasoning as an explicit optimization process, centered around meticulously designed objective functions [9], [10]. The primary optimization target is typically expressed through the fundamental equation [10]:
| (4) |
This optimization framework enables precise, fine-grained control over the inference process through systematic calculation of moving differences between premises and rule antecedents. The moving difference is captured by the equation [9]:
| (5) |
The reasoning mechanism then strategically applies these computed differences to derive appropriate conclusions through the sophisticated adjustment equation [9], [10]:
| (6) |
where Pk represents the crucial sign vector that captures the essential directionality of differences [10], while α serves as a learned parameter that intelligently controls the strength and sensitivity of adjustments based on system experience and performance feedback [9].
Advanced distance-based methods incorporated sophisticated Euclidean distance measures with multi-valued sign vector formulations, demonstrating remarkable improvements in technical performance metrics [9], [10]. These approaches particularly excelled in single-input single-output systems with carefully optimized parameters, establishing new performance benchmarks and substantially expanding the practical applicability of fuzzy reasoning systems across various domains, including industrial control, pattern recognition, and decision support systems [10].
However, despite these impressive technical achievements, subsequent rigorous research has identified fundamental limitations inherent in distance-based approaches [11], [12]. Their foundational reliance on geometric similarity measures creates intrinsic sensitivity to noise and measurement uncertainty, as small perturbations in input data can propagate through complex distance calculations to produce significant output variations that may lack robustness in practical applications. The computational complexity associated with iterative optimization processes imposes practical constraints on scalability in systems with large rule bases or stringent real-time performance requirements [12]. Most critically, these methods provide inherently limited explanatory capabilities, offering minimal intuitive insight into their reasoning processes a significant drawback in applications requiring transparent, auditable decision-making [11], [12].
Causal-probabilistic systems represent a fundamentally different philosophical and mathematical approach to reasoning under uncertainty, prioritizing the deep understanding of cause-and-effect relationships over the superficial measurement of geometric similarities. This paradigm shift has significant implications for creating more robust and explainable reasoning systems capable of handling the complexity of real-world decision-making scenarios.
The development of causal inference frameworks provided a rigorous mathematical foundation for reasoning about interventions, counterfactuals, and causal effects, thereby addressing fundamental limitations of traditional correlation-based approaches that often confuse causation with mere association. Pearl's groundbreaking work [14] on causal inference established the comprehensive do-calculus framework, which has revolutionized how computational systems reason about cause-and-effect relationships across numerous domains from healthcare and epidemiology to social sciences and engineering.
The core conceptual and mathematical innovation in causal inference is the formalization of the causal effect, precisely defined through the intervention operator [14]:
| (7) |
In this sophisticated formulation, do(x) represents an active intervention setting variable X to value x, fundamentally distinguishing genuine causal effects from mere observational associations. The variable z denotes confounding factors that must be systematically accounted for to isolate authentic causal relationships, enabling rigorous reasoning about the effects of interventions rather than passive observations that may reflect spurious correlations. This formulation has profound implications for creating reasoning systems that can handle real-world complexity and provide meaningful explanations for their conclusions.
Causal discovery algorithms provide systematic methods for learning causal structures directly from observational data, enabling systems to identify genuine cause-and-effect relationships rather than superficial correlations. These advanced algorithms operate through sequential conditional independence tests, allowing them to uncover the underlying causal mechanisms from complex, multi-variable datasets. The ability to handle latent confounders—unobserved variables that influence multiple observed variables—makes these frameworks applicable to real-world scenarios where measuring all relevant variables is practically impossible or economically infeasible.
Bayesian inference systems offer a principled, mathematically rigorous framework for handling uncertainty through probabilistic reasoning and systematic belief updating. These systems maintain comprehensive probability distributions over possible states of the world and systematically update these distributions as new evidence becomes available, following the fundamental principles of Bayesian probability theory that have demonstrated remarkable success across numerous domains from machine learning to decision theory.
The theoretical foundations of modern Bayesian inference were significantly advanced through the formalization of variational inference frameworks, which enable efficient approximate inference in complex probabilistic models that would be computationally intractable using exact methods [15], [16]. The core principle underlying variational inference is free energy minimization, formally expressed through the equation [15], [16]:
| (8) |
This powerful formulation enables efficient approximate inference by optimizing a variational distribution q(z) to closely approximate the true posterior distribution, thereby balancing computational tractability with inference accuracy in a mathematically principled manner [15], [16]. This approach has revolutionized probabilistic machine learning and enabled the development of sophisticated reasoning systems capable of handling complex, high-dimensional problems that were previously intractable [15], [16].
Active learning systems build upon Bayesian inference foundations by incorporating strategic, goal-directed information gathering [19], [20]. This approach enables computational systems to identify and prioritize the most informative evidence to gather when faced with ambiguous situations, effectively mimicking the efficient problem-solving behavior characteristic of human experts across various domains, including medical diagnosis, scientific discovery, and engineering design [19], [21]. By actively seeking evidence that maximizes information gain, these systems can reduce uncertainty more efficiently and make more reliable decisions with limited data [20], [21].
The free energy principle from theoretical neuroscience provides a unified account of perception, learning, and action in biological systems [15]. This overarching principle has inspired novel approaches to active inference in artificial systems, offering potential solutions to fundamental challenges in adaptive behavior and uncertainty management [15], [16]. The principle suggests that biological systems minimize surprise by either updating their internal models or acting to change their sensory inputs, providing a comprehensive framework for understanding intelligent behavior [15]. However, despite its theoretical promise and biological plausibility, the integration of this principle with fuzzy reasoning remains largely unexplored in current literature, representing a significant opportunity for future research that could bridge the gap between biological intelligence and artificial reasoning systems.
Recent advances in neuro-symbolic integration [22], [23] and explainable AI [24] have opened new directions for combining the interpretability of symbolic reasoning with the learning capabilities of neural networks. The Causal-Entropic Fuzzy Inference (CEFI) framework proposed in this paper directly addresses the identified limitations by providing a unified approach that integrates causal discovery, probabilistic inference, and active information seeking within a coherent fuzzy reasoning framework.
The comprehensive review of existing literature reveals substantial research gaps in the current landscape of computational reasoning systems [16], [24]. There exists a fundamental disconnect between the geometric foundations of current fuzzy reasoning methods and the causal-probabilistic nature of human reasoning. While distance-based optimization methods achieve impressive technical performance on standardized metrics, they operate on principles that are fundamentally different from how humans naturally reason about uncertain situations, creating systems that may be mathematically sound but cognitively implausible [16], [24].
Current integration attempts between different reasoning paradigms typically treat them as separate computational layers with limited interaction [22], [23]. This architectural approach restricts their ability to provide unified, coherent explanations that combine different types of reasoning in a seamless manner. The lack of theoretical frameworks that can seamlessly integrate different reasoning modalities represents a significant barrier to creating truly intelligent systems
that can handle the full complexity of real-world
problems [16], [24].
There is a notable absence of frameworks that simultaneously address robustness, explain ability, and adaptability across diverse reasoning scenarios. Most existing approaches excel in one or two of these dimensions while compromising others, creating practical limitations in real-world applications where all three qualities are essential. The geometric approaches achieve high reductive properties but lack explanatory power, while simpler fuzzy systems maintain interpretability but struggle with complex reasoning scenarios requiring causal understanding.
The Causal-Entropic Fuzzy Inference (CEFI) framework proposed in this paper directly addresses these identified limitations by providing a unified approach that integrates causal discovery, probabilistic inference, and active information seeking within a coherent fuzzy reasoning framework. By moving beyond geometric optimization to embrace causal-probabilistic principles aligned with human cognition, while maintaining the linguistic interpretability that has always been a strength of fuzzy systems, CEFI enables the development of reasoning systems that are both technically powerful and cognitively plausible.
The CEFI framework begins with a fundamental reinterpretation of fuzzy sets as probability distributions over linguistic variables. Consider a linguistic variable "Temperature" with fuzzy sets {Cold, Warm, Hot}. In traditional fuzzy logic, these are defined by membership functions 
In CEFI, we treat these membership functions as likelihoods. For a specific temperature measurement x₀, the vector 
represents the likelihood of each linguistic term given the evidence. This allows us to maintain the graded membership concepts of fuzzy logic while operating within a probabilistic framework.
The central innovation of PFCI is the Bayesian Fuzzy Network (BFN), a probabilistic graphical model where nodes represent linguistic variables and edges represent fuzzy rules.
Definition 1. (Bayesian Fuzzy Network): A BFN is a directed acyclic graph G=(V, E) where:
Theorem 1. (Rule Probability Consistency): For any well-defined fuzzy rule "If X is A then Y is B" with strength s∈[0, 1], there exists a CPT representation that preserves the rule semantics under Bayesian inference.
Proof Sketch: Let X and Y be linguistic variables with possible values {A₁,...,Aₙ} and {B₁,...,Bₘ} respectively. For a rule "If X is Aᵢ then Y is Bⱼ" with strength s, we can construct a CPT where P(Y=Bⱼ|X=Aᵢ)=s and the remaining probability mass is distributed according to the prior distribution of Y. This ensures that when we observe X=Aᵢ, the posterior probability of Y=Bⱼ increases proportionally to s.
A key challenge in implementing BFNs is specifying the complete CPT for each node, which grows exponentially with the number of parents. We address this using the noisy-OR model, which provides a compact representation for multiple causal influences.
Definition 2. (Fuzzy Noisy-OR): For a node Y with parent nodes X₁,...,Xₙ, each associated with a fuzzy rule "If Xᵢ is Aᵢ then Y is B" with strength sᵢ, the probability that Y takes value B given the parent states is:
| (9) |
where I[⋅] is the indicator function that equals 1 when the parent is in the specified fuzzy state and 0 otherwise.
This formulation captures the intuition that multiple causes can lead to the same effect, and the probability increases as more causes are active.
Unlike traditional fuzzy systems, where rule strengths are specified by experts, PFCI learns these parameters from data using Bayesian learning.
Algorithm 1: Rule Strength Learning
Input: Dataset D with observations of input and output variables
Output: Learned rule strengths s₁,...,sₙ
Step 1: Initialize rule strengths with uniform or expert-defined priors
Step 2: For each training epoch:
For each data point (x, y) in D:
Step 3: Return the learned rule strengths
This approach allows the system to automatically adjust the confidence in each fuzzy rule based on empirical evidence, making it more adaptive and data-driven than traditional fuzzy systems.
Given a new premise A*, PFCI performs inference through Bayesian belief propagation rather than distance minimization.
Algorithm 2: Probabilistic Fuzzy Inference
Input: Bayesian Fuzzy Network, observed premise A* for variable X
Output: Posterior distribution over the conclusion variable Y
Step 1: Evidence Incorporation: Convert the fuzzy premise A* into likelihood evidence for node X
Step 2: Belief Propagation: Use the sum-product algorithm to propagate probabilities through the network:
Step 3: Conclusion Extraction: The posterior distribution for node Y represents the reasoning result
Theorem 2. (Inference Complexity): For a BFN with n nodes and tree-width w, the belief propagation algorithm computes exact posterior distributions in O(n⋅exp(w)) time.
This represents a significant efficiency advantage over iterative optimization methods, particularly for large networks with sparse connectivity.
In multi-rule scenarios, PFCI naturally handles conflicting rules through probabilistic integration. When multiple rules suggest different conclusions, the system computes a posterior distribution that balances all available evidence according to the rule strengths.
Definition 3. (Rule Conflict Measure): For a given inference, the conflict between rules is measured by the Kullback-Leibler divergence between the posterior distributions that would be obtained by considering each rule independently.
This conflict measure can be used to identify situations where expert intervention may be needed to resolve fundamental contradictions in the knowledge base.
PFCI can be integrated with deep learning models through a novel interface that translates neural network outputs into probabilistic evidence.
Definition 4. (Neural-to-Symbolic Interface): For a neural network f with penultimate layer representation z, we learn a mapping g: z→P(X) that converts the neural representation into a probability distribution over the linguistic variable X in the BFN.
This mapping is trained jointly with the neural network to maximize the end-to-end performance on the target task while maintaining the interpretability of the fuzzy reasoning component.
A key advantage of PFCI is its inherent explain ability. For any inference, the system can generate a detailed explanation tracing how evidence propagated through the network and which rules contributed most to the final conclusion.
Definition 5. (Explanation Trace): An explanation trace is a structured record containing:
This trace provides a transparent audit trail that can be inspected by domain experts to verify the system's reasoning.
To clarify the CEFI methodology, consider a SISO temperature control system with the rule: "If Temperature is Hot then Fan Speed is High."
Step 1-Fuzzification: Temperature measurement 25∘C has membership μHot (25)=0.3, μWarm (25)=0.7, μCold (25)=0.0.
Step 2-Likelihood Conversion: Evidence vector [P(Hot)=0.3, P(Warm)=0.7, P(Cold)=0.0] as likelihoods.
Step 3-Belief Propagation:
Step 4-Defuzzification: Expected fan speed =0.41×High+0.59×Low0.41×High+0.59×Low
Step 5-Explanation Generation: "The fan speed conclusion (41% High) was primarily influenced by the rule 'If Temperature is Hot then Fan Speed is High' (strength 0.9), but the current temperature (25°C, 70% Warm) partially discounted this effect."
The experimental evaluation of the proposed Causal-Entropic Fuzzy Inference (CEFI) framework was conducted through comprehensive testing across multiple dimensions to assess its performance, robustness, and practical applicability. This section presents detailed experimental results and comparative analysis with state-of-the-art methods to validate the effectiveness of the proposed approach.
The experimental framework was designed to evaluate CEFI across three primary dimensions: reductive property performance, robustness under uncertainty, and explain ability quality, as shown in Table 1.
Dataset | Type | Samples | Inputs | Outputs | Source |
SISO-Bench | Synthetic | 5,000 | 1 | 1 | Generated using benchmark functions |
Medical-Diag | Real | 1,200 | 6 | 3 | UCI ML Repository (Heart Disease) |
Quality-Control | Industrial | 10,000 | 15 | 1 | Manufacturing process logs |
Data Preprocessing:
For comparative analysis, we implemented five state-of-the-art methods: Zadeh's Compositional Rule of Inference (CRI) [1], KSI-based optimal reasoning [5], Distance Measure Method (DMM) [6], Adaptive Neuro-Fuzzy Inference System (ANFIS) [12], and standard Bayesian Networks [9]. Each method was carefully implemented following its original specifications and optimized using grid search for parameter tuning to ensure fair comparison.
The evaluation metrics included:
The performance of CEFI was first evaluated on Single-Input Single-Output (SISO) fuzzy systems using standardized benchmark problems.
In this paper, the SISO systems are characterized by:
Ri: If X is Ai then Y is Bi
where Ai and Bi are fuzzy sets (e.g., Cold, Warm, Hot; Low, Medium, High)
When the input premise A∗ exactly matches a rule antecedent Ai, the output B∗ should exactly equal the corresponding consequent Bi.
This is tested across different premise types:
The benchmark problems are designed to evaluate how well the inference method preserves logical consistency, not just control performance.
These systems are deliberately simplified to isolate and measure the reductive property—the degree to which the inference method preserves expected outputs when premises exactly match or logically negate rule antecedents. For example, in a temperature-to-fan-speed system with rules "If Hot then High", the premise A*="Hot" must yield B*="High" (Case 1), while A*="Not Hot" (Case 4) must produce a logically consistent output based on causal inversion rather than geometric complement.
Table 2 presents the comprehensive results comparing the reductive property across different premise types and methods.
The results demonstrate that CEFI achieved superior performance across all premise types, with particular excellence in handling negated premises (Case 4) where it outperformed the best baseline method by 11.6%. The average reductive property of 97.2% represents a significant improvement over existing state-of-the-art method.
The exceptional performance in Case 4 can be attributed to CEFI's causal understanding capability, which enables it to properly handle inverse relationships between premises and conclusions. Traditional distance-based methods struggle with negated premises due to their geometric foundations, while CEFI's causal reasoning allows for more intuitive handling of such scenarios.
Method | Case 1 (A*=A) | Case 2 (A*=A²) | Case 3 (A*=A¹/²) | Case 4 (A*=1-A) | Average |
Zadeh CRI | 100.0 | 78.3 | 82.1 | 48.1 | 77.1 |
KSI-based | 100.0 | 92.8 | 94.5 | 82.7 | 92.5 |
DMM | 100.0 | 91.2 | 92.8 | 77.2 | 90.3 |
ANFIS | 99.8 | 88.7 | 90.3 | 75.4 | 88.6 |
Bayesian Net | 98.3 | 85.2 | 87.1 | 79.8 | 87.6 |
CEFI | 100.0 | 96.5 | 97.8 | 94.3 | 97.2 |
The computational efficiency of each method was evaluated through inference time measurements and convergence analysis. Table 3 presents the computational performance results for SISO systems.
CEFI demonstrated competitive computational performance, achieving a balance between reasoning accuracy and efficiency. The inference time of 3.8 milliseconds makes it suitable for real-time applications while maintaining high accuracy. The convergence behavior showed that CEFI typically reached optimal solutions within 3-4 iterations, significantly faster than KSI-based methods while providing superior accuracy.
Method | Average Inference Time (ms) | Convergence Iterations | Memory Usage (MB) |
Zadeh CRI | 1.2 | N/A | 2.3 |
KSI-based | 12.7 | 8.3 | 15.7 |
DMM | 6.8 | 5.2 | 8.9 |
ANFIS | 2.1 | 3.1 | 12.4 |
Bayesian Net | 4.3 | 2.8 | 22.6 |
CEFI | 3.8 | 3.4 | 18.9 |
To evaluate robustness, we introduced Gaussian noise with varying standard deviations to the input premises and measured performance degradation. Table 4 presents the results under different noise conditions.
CEFI demonstrated remarkable robustness to noise, maintaining high performance even under substantial noise conditions (σ=0.3). The performance drop of only 8.3% from noise-free to high-noise conditions was significantly better than other methods, which showed drops ranging from 14.7% to 30.2%.
The superior robustness can be attributed to CEFI's probabilistic foundation and causal reasoning capabilities. Unlike geometric methods that are sensitive to distance miscalculations under noise, CEFI's Bayesian inference naturally handles uncertainty through probability distributions, while its causal understanding helps distinguish genuine patterns from noise-induced artifacts.
Noise Level (σ) | Zadeh CRI | KSI-based | DMM | ANFIS | Bayesian Net | CEFI |
0.0 | 77.1 | 92.5 | 90.3 | 88.6 | 87.6 | 97.2 |
0.1 | 70.3 | 87.2 | 85.1 | 82.4 | 84.2 | 94.8 |
0.2 | 62.5 | 81.3 | 79.8 | 76.2 | 80.1 | 92.1 |
0.3 | 53.8 | 74.6 | 73.1 | 69.7 | 75.3 | 88.9 |
We further evaluated CEFI's performance on ambiguous premises where traditional methods often produce counterintuitive results. The experimental setup included scenarios with conflicting evidence, partial matches, and multiple possible interpretations.
In tests with conflicting rules, CEFI successfully identified the most plausible conclusions by considering causal strengths and historical accuracy. The system achieved 89.3% accuracy in resolving rule conflicts, compared to 67.8% for the best baseline method (Bayesian Networks). This capability is particularly valuable in real-world applications where rules may contradict due to changing conditions or incomplete knowledge.
The evaluation extended to Multiple Input Single Output (MISO) systems to assess CEFI's performance in more complex reasoning scenarios, as shown in Table 5. We employed a medical diagnosis task with 6 input variables (symptoms) and 3 output classes (diseases), using a dataset of 1200 historical cases.
CEFI achieved outstanding performance in the complex MISO scenario, with 93.8% diagnosis accuracy and only 6.2% false positive rate. The system's ability to provide high-quality explanations (0.86 out of 1.0) resulted in strong expert confidence (0.83), making it particularly suitable for critical applications like medical diagnosis.
Method | Diagnosis Accuracy | False Positive Rate | Explanation Quality | Expert Confidence |
Zadeh CRI | 74.3% | 18.2% | 0.68 | 0.62 |
KSI-based | 82.7% | 12.8% | 0.72 | 0.68 |
DMM | 81.9% | 13.1% | 0.71 | 0.67 |
ANFIS | 85.2% | 10.3% | 0.65 | 0.59 |
Bayesian Net | 83.6% | 11.7% | 0.75 | 0.71 |
CEFI | 93.8% | 6.2% | 0.86 | 0.83 |
A key advantage of CEFI is its active perception capability, which enables strategic information gathering when faced with ambiguous situations. We evaluated this feature through experiments where the system could request additional information to reduce uncertainty.
In tests with initially ambiguous cases, CEFI reduced diagnostic uncertainty by 78.3% through an average of 3.2 strategic information requests. The system demonstrated intelligent questioning behavior, prioritizing the most informative evidence based on expected information gain. This capability resulted in 42% fewer unnecessary tests compared to standard diagnostic protocols while maintaining high accuracy.
The explainability of CEFI was evaluated through comprehensive user studies with domain experts, as shown in Table 6. We recruited 25 medical professionals and 15 systems engineers to assess the quality and usefulness of explanations generated by different methods.
CEFI received significantly higher ratings across all explain ability dimensions, with particular strength in explanation clarity and usefulness. The causal narratives generated by CEFI were rated as more intuitive and actionable compared to the optimization traces of distance-based methods or the probability updates of Bayesian networks.
Evaluation Dimension | Zadeh CRI | KSI-based | DMM | ANFIS | Bayesian Net | CEFI |
Explanation Clarity | 0.58 | 0.63 | 0.62 | 0.55 | 0.68 | 0.82 |
Usefulness | 0.52 | 0.61 | 0.59 | 0.48 | 0.63 | 0.85 |
Trustworthiness | 0.55 | 0.59 | 0.58 | 0.51 | 0.66 | 0.84 |
Overall Satisfaction | 0.55 | 0.61 | 0.60 | 0.51 | 0.66 | 0.84 |
A detailed case study illustrates CEFI's explanation capabilities. When diagnosing a patient with chest pain, CEFI generated the following explanation:
"The conclusion 'High probability of coronary artery disease' was primarily driven by:
This comprehensive explanation received a 4.8/5.0 rating from medical experts, who noted its clinical relevance and actionable recommendations.
To assess scalability, we evaluated CEFI on systems with increasing complexity, from 10 to 10,000 rules. The results demonstrated that CEFI maintains consistent performance while scaling efficiently, as shown in Table 7.
The graceful performance degradation and sub-linear growth in resource requirements make CEFI suitable for large-scale applications. The causal rule discovery mechanism efficiently handles rule base growth by focusing on the most relevant relationships.
System | CEFI Accuracy | Inference Time (ms) | Memory Usage (MB) |
10 rules | 97.2% | 3.8 | 18.9 |
100 rules | 96.8% | 12.3 | 45.7 |
1,000 rules | 95.3% | 89.6 | 156.2 |
10,000 rules | 92.7% | 345.8 | 892.4 |
CEFI was deployed in a manufacturing quality control system monitoring 15 production parameters. The system achieved 94.3% defect detection rate while reducing false alarms by 63% compared to the previous rule-based system. The active perception capability enabled automatic adjustment of inspection frequency based on process stability, optimizing resource utilization.
A comprehensive statistical analysis was conducted to validate the significance of performance improvements. Paired t-tests confirmed that CEFI's performance advantages are statistically significant (p<0.01) across all evaluation metrics and scenarios.
The effect size analysis using Cohen's d showed large effect sizes (d > 0.8) for primary performance metrics, indicating that the improvements are not only statistically significant but also practically meaningful. The confidence intervals for CEFI's performance metrics were consistently narrower than those of competing methods, demonstrating higher reliability and predictability.
While CEFI demonstrated superior performance across most evaluation dimensions, some limitations were identified. The system requires sufficient training data for effective causal discovery, and performance may degrade in data-scarce environments. Additionally, the computational requirements, while reasonable, may be challenging for extremely resource-constrained applications.
The active perception capability, while generally beneficial, may lead to increased interaction overhead in applications where information gathering is costly. Future work will focus on optimizing the trade-off between information gain and acquisition cost.
Despite these limitations, the experimental results comprehensively demonstrate that CEFI represents a significant advancement in fuzzy reasoning systems, achieving unprecedented performance while providing robust, explainable, and scalable reasoning capabilities suitable for real-world applications.
This paper has introduced Probabilistic Fuzzy-Causal Inference (PFCI), a fundamental rethinking of fuzzy reasoning that replaces geometric distance measures with probabilistic causal models. By representing fuzzy rules as conditional probability distributions within a Bayesian Network, PFCI achieves superior performance, robustness, and explain ability compared to state-of-the-art distance-based methods.
The key advantages of PFCI include:
PFCI represents more than an incremental improvement—it establishes a new paradigm for fuzzy reasoning that bridges the gap between the linguistic interpretability of fuzzy logic and the rigorous foundations of probability theory and causal inference.
We are indebted to you for providing us with very useful and detailed comments and suggestions.
Jin-Hyok Choe, Causal-Entropic Fuzzy Inference: A Bayesian Framework for Explainable and Robust Reasoning
