Abstract
In this paper, an immune inspired multi-objective fuzzy modeling (IMOFM) mechanism is proposed specifically for high-dimensional regression problems. For such problems, prediction accuracy is often the paramount requirement. With such a requirement in mind, however, one should also put considerable efforts in eliciting models which are as transparent as possible, a 'tricky' exercise in itself. The proposed mechanism adopts a multistage modeling procedure and a variable length coding scheme to account for the enlarged search space due to simultaneous optimisation of the rule-base structure and its associated parameters. We claim here that IMOFM can account for both Singleton and Mamdani Fuzzy Rule-Based Systems (FRBS) due to the carefully chosen output membership functions, the inference scheme and the defuzzification method. The proposed modeling approach has been compared to other representatives using a benchmark problem, and was further applied to a high-dimensional problem, taken from the steel industry, which concerns the prediction of mechanical properties of hot rolled steels. Results confirm that IMOFM is capable of eliciting not only accurate but also transparent FRBSs from quantitative data.
Introduction
In data-driven modeling, researchers frequently encounter the "Principle of Incompatibility": as the precision of a model increases, its interpretability—or transparency—typically decreases. Fuzzy Rule-Based Systems (FRBS) are often favored for their ability to represent complex systems through human-readable "If-Then" logic. However, when applied to high-dimensional datasets, these systems frequently fall victim to the "curse of dimensionality," resulting in hundreds of overlapping rules that obscure the underlying logic.
To address this, the Immune-inspired Multi-Objective Fuzzy Modeling (IMOFM) framework was developed. It aims to bridge the gap between predictive power and model simplicity, ensuring that even complex industrial models remain accessible to human experts.
The IMOFM Framework
The core of the IMOFM approach is a structured three-stage modeling procedure designed to evolve a rough data draft into a polished, optimized system.
- Initial Extraction: The process begins by using an evolutionary K-means clustering algorithm to generate an initial set of fuzzy rules directly from raw data.
- Refinement: This "first draft" is then fine-tuned using a constrained Back-Error-Propagation (BEP) algorithm. This stage acts as a safety net; by polishing the initial model, the system ensures that high-accuracy candidates are not prematurely discarded during the subsequent competitive optimization phase.
- Multi-Objective Optimization: The final stage utilizes a Population Adaptive Immune Algorithm (PAIA) to evolve a set of Pareto-optimal models.
Technical Innovations
A key technical hurdle in multi-objective fuzzy modeling is the search space. IMOFM employs Variable Length Coding, allowing the algorithm to dynamically add or remove rules without the search space exploding. Furthermore, the PAIA mimics biological immune responses, where "affinity" measures the balance between low error and low complexity. During this process, the system performs automated simplification, merging similar fuzzy sets and deleting redundant rules that do not contribute to the model's predictive accuracy.
Balancing Accuracy and Complexity
The multi-objective nature of the third stage allows for the generation of a "Pareto Front." This front represents a set of non-dominated solutions where no single model can improve its accuracy without increasing its complexity (number of rules), and vice versa.
By visualizing this trade-off, practitioners can select a "Final Solution" that provides the necessary accuracy for the task while remaining compact enough for human verification. This is particularly vital in industrial settings where understanding the why behind a prediction is as important as the prediction itself.
Experimental Results
The framework was validated against both a standard nonlinear benchmark and a high-dimensional real-world dataset from the steel industry.
- Nonlinear Benchmark: IMOFM achieved a Root Mean Square Error (RMSE) of approximately 0.068 using only 5 rules. This outperformed several established methods that either required more complex rule bases or yielded higher error rates.
- Steel Industry Data (UTS): The framework was applied to predict the Ultimate Tensile Strength (UTS) of hot-rolled steels, a problem involving 15 input variables and over 3,700 samples. IMOFM successfully reduced the RMSE from an initial ~120 to ~31.2. More impressively, it condensed the model into a handful of interpretable rules (typically 6 to 10), which metallurgists confirmed aligned with physical principles.
Conclusion
IMOFM demonstrates that high-dimensional regression does not necessitate a "black-box" approach. By integrating a three-stage refinement process with an immune-inspired optimization algorithm, the framework produces fuzzy models that are both accurate and transparent. Future work will focus on reducing the computational cost of the multi-objective optimization and extending the framework to more complex industrial control loops.
Summary updated