Abstract
This paper discusses how to construct a fuzzy classification system (FCS) effectively and accurately. In the FCS, the initial fuzzy rules are optimized with a quantum bit which has many unique advantages such as small population size, fast convergence, short training time and strong global search ability. After then, in order to accomplish the optimization for the fuzzy space partition and the number of fuzzy rules, this paper propose a method-quantum particle swarm optimization (QPSO) -to improve the initial FCS. The experiment result demonstrates that this method is more efficient than other methods without QPSO.
Introduction
Fuzzy Classification Systems (FCS) are prized for their interpretability, utilizing human-readable "IF-THEN" rules to make decisions. However, building an effective FCS is a balancing act: excessive rules lead to the "curse of dimensionality," while too few rules compromise accuracy. Traditionally, these rules are hand-crafted or evolved using standard Genetic Algorithms (GA), which often struggle with local optima and require large populations. This research introduces a hybrid approach using Quantum-inspired Particle Swarm Optimization (QPSO) to optimize both fuzzy space partitioning and rule sets simultaneously, achieving high accuracy with a minimal number of rules.
Methodology
The proposed system integrates Fuzzy Logic with QPSO through several key innovations:
- Quantum Bit (Qubit) Representation: Instead of standard binary bits, the system uses qubits to represent the probability of a feature being selected or a rule being active. This allows for a "superposition" of states, granting the algorithm a much stronger global search ability with a smaller population size.
- QPSO Algorithm: Unlike standard PSO, which uses velocity and position vectors, QPSO treats particles as having quantum-like behavior. Particles update their positions based on a "mean best" position of the whole swarm, governed by a contraction-expansion coefficient.
- Rotation Gates: To evolve the population, the algorithm uses quantum rotation gates. These mathematically "rotate" the probability vectors toward the best-known solutions, ensuring fast convergence.
- Multi-Objective Fitness Function: The algorithm's fitness function rewards high classification accuracy while specifically penalizing a high number of fuzzy rules. This forces the system to "prune" unnecessary complexity.
Experimental Results
The researchers evaluated their method using classic benchmarking datasets, primarily the Iris and Wine datasets.
Iris Dataset Performance
The Iris dataset consists of 4 dimensions and 150 samples across 3 species. The QPSO-enhanced system significantly optimized the initial model, reducing the rule set from 16 to 9 rules while improving accuracy from 92.45% to 96.67%.
Wine Dataset Performance
The Wine dataset is more complex, featuring 13 attributes and 178 samples. On this dataset, the QPSO method demonstrated even more dramatic efficiency, slashing the rule count from 32 down to just 10, while boosting accuracy from 87.87% to 91.57%.
Discussion
A key insight from the study was the discovery of "dummy rules"—invalid rules covering areas of the input space where no data actually existed. The QPSO was highly effective at identifying and eliminating these redundancies. Furthermore, by using quantum-inspired mathematics, the researchers achieved superior results with a population of only 20 individuals over 100 generations. This efficiency demonstrates that interpretability and accuracy are not mutually exclusive; by optimizing rule antecedents through QPSO, the system finds an elegant way to divide the data space that boosts performance while cutting complexity.
Conclusion and Future Work
The QPSO-enhanced FCS consistently outperformed both direct fuzzy modeling and standard GA approaches. While the results on the Iris and Wine datasets are promising, future work should address the computational complexity of quantum rotation gate math versus standard operations. Additionally, exploring the scalability of this approach to high-dimensional "Big Data" scenarios remains a priority. The ultimate goal is to further simplify model structures to create compact, highly accurate rule bases for real-time applications.
Summary updated