ANFIS-PSO Hybrid Optimization

• ANFIS-PSO is a hybrid model that optimizes fuzzy inference system parameters using PSO's global search to overcome local minima.
• It employs a five-layer ANFIS structure with parameterized membership functions to effectively model nonlinear systems.
• The approach has been successfully applied in industrial control, biomedical signal processing, and environmental prediction with improved error metrics.

A Particle Swarm Optimization–based Adaptive Neuro-Fuzzy Inference System (ANFIS-PSO) is a hybrid computational intelligence framework in which the parameters of an ANFIS model are globally optimized via a Particle Swarm Optimization (PSO) algorithm. This synergistic integration exploits the universal approximation properties and interpretability of ANFIS while ameliorating the local-minima limitations of conventional gradient-based or hybrid learning methods through PSO's population-based stochastic global search. Applications of ANFIS-PSO span regression and classification tasks in industrial control, biomedical signal processing, environmental prediction, and model predictive control (MPC) adaptation.

1. ANFIS Model Structure and Parameterization

An ANFIS implements a Takagi–Sugeno fuzzy inference system in a five-layer feedforward architecture, with the following canonical structure:

Inputs and Fuzzification:

Each input $x_j$ is fuzzified by a set of parameterized membership functions (MFs), commonly Gaussian: $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$ or generalized bell: $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$ where MF parameters $(c, \sigma)$ or $(a, b, c)$ are the premise parameters.

Rule Layer:

With $m$ MFs per input and $d$ inputs, the full rule base contains $m^d$ rules. Rule firing strength for rule $i$: $\omega_i = \prod_{j=1}^{d} \mu_{A_{ij}}(x_j)$

Normalization:

$$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$0

where $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$1.

Consequent Layer:

For first-order Sugeno (TS) models, each rule's output is a linear combination: $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$2 with consequent parameters $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$3.

Output Aggregation:

$$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$4

Parameter space dimensionality is high, especially for larger $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$5, motivating the adoption of global search methods for parameter tuning.

2. Particle Swarm Optimization for ANFIS Tuning

PSO encodes all tunable parameters—premise ($$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$6 or $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$7) and consequent ($$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$8)—within the position vector of each particle. Standard PSO is applied as follows: $$\mu_A(x) = \exp\left(-\frac{(x-c)^2}{2\sigma^2}\right)$$9

$$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$0

where $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$1 is particle velocity, $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$2 is position (ANFIS parameterization), $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$3 is particle's best, $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$4 is global best, $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$5, and $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$6 are PSO hyperparameters.

Problem-dependent variants include exponential decay of inertia weight: $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$7 and adaptive adjustment of cognitive/social coefficients. The fitness function is task-dependent (MSE, RMSE, accuracy, or domain-specific metrics).

3. Workflow and Algorithmic Integration

The ANFIS-PSO optimization loop replaces or augments classical ANFIS learning as follows:

1. Initialization: Define ANFIS structure (number/type of MFs per input, rule base, etc.). Randomly initialize the PSO swarm in the high-dimensional parameter space.
2. Fitness Evaluation: For each particle, decode its position into ANFIS parameters; compute task-specific error on training or validation data.
3. PSO Update: Update particle velocities and positions. Apply parameter bounds; enforce MF constraints (e.g., positive widths).
4. Convergence and Model Selection: Terminate upon reaching maximum iterations, stagnating improvement, or error threshold.
5. Finalization: The global best particle at end of optimization yields the tuned ANFIS model.

This approach is systematically applied both as a standalone model (e.g., for regression/classification of engineered or biomedical systems) and embedded within larger control architectures (e.g., adaptive MPC).

4. Application Domains and Representative Results

ANFIS-PSO is widely adopted across diverse application areas, with task-specific architectures and metrics:

• MPC Adaptation for Autonomous Vehicle Path Tracking: ANFIS-PSO tuned networks map longitudinal speed, wind speed, adhesion, and reference position to MPC hyperparameters ($$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$8). Offline PSO finds optimal ANFIS training data, followed by hybrid learning. In triple lane-change scenarios: standard MPC yields MSE ≃ 0.0318, ANFIS-MPC ≃ 0.0062, with NN-MPC slightly more accurate and ANFIS-MPC yielding smoother steering (Kebbati et al., 21 Sep 2025).
• Motor Imagery EEG Classification: ANFIS, fed by filter-bank CSP features, is globally tuned by PSO for all MF and rule-consequent parameters. Within-subject accuracy: ANFIS–FBCSP–PSO 68.58 ± 13.76%, higher than deep learning baseline EEGNet (63.79 ± 8.49%) (Aktar et al., 1 Nov 2025).
• Diagnostic and Process Regression:
  • Liver disease classification: PSO-tuned ANFIS achieves 10% higher accuracy and lower RMSE over vanilla ANFIS (Rajabi et al., 2019).
  • HVAC exergy prediction: ANFIS-PSO RMSE = 0.0065, MAE = 0.0028, $$\mu_A(x) = \frac{1}{1+\left|\frac{x-c}{a}\right|^{2b}}$$9, outperforming both ANFIS-GA and standalone ANFIS (Ardabili et al., 2020).
  • Mercury emission modeling: ANFIS-PSO attains $(c, \sigma)$0, MARE% = 0.0133 on test data, far surpassing hybrid ANN or SVM models (Shamshirband et al., 2019).
• Engineering Systems: Prediction of air velocity in dam bottom outlets, exergy prediction, power plant emissions, etc., with ANFIS-PSO yielding lower error and higher correlation than competing neuroevolutionary or machine learning methods (Narimani et al., 2021).

5. Comparative Advantages and Limitations

Advantages:

• Global optimization across highly multimodal parameter landscapes.
• Joint tuning of both nonlinear premise and linear consequent parameters, bypassing the local minima and premature convergence associated with hybrid learning or gradient descent.
• Effective in moderate-to-high dimensional ANFIS parameterizations, subject to computational resource constraints.
• Consistently yields lower error metrics (RMSE/MSE/MARE) and higher measures of fit ($(c, \sigma)$1, accuracy, κ) relative to both untuned ANFIS and non-fuzzy ML baselines.

Limitations:

• Rapid search space growth with increasing number of MFs and inputs: for $(c, \sigma)$2 MFs and $(c, \sigma)$3 inputs, parameter dimension $(c, \sigma)$4. Computational cost may become prohibitive for large architectures.
• PSO hyperparameters (swarm size, inertia, cognitive/social factors) require careful selection; improper tuning can result in poor convergence or overfitting.
• No guarantee of interpretability preservation if an excessively large rule base is used; "curse of dimensionality" can impact both computational tractability and qualitative interpretability (Kebbati et al., 21 Sep 2025).

6. Implementation Details and Guidelines

Empirical studies employ swarms of 20–250 particles, 50–400 iterations, and commonly use linearly or exponentially decreasing inertia ($(c, \sigma)$5) with $(c, \sigma)$6 in [1.5, 2.0]. Standardization of data, appropriate bounding of MF parameters, and careful selection of MF type (triangular, Gaussian, generalized bell) are necessary for numerical stability and meaningful models (Narimani et al., 2021).

For real-time and resource-constrained settings, moderate values (MFs per input, rules per network) are recommended, e.g., 8–16 rules for EEG classification, 3 MFs per input for industrial regression (Aktar et al., 1 Nov 2025, Ardabili et al., 2020). Parameter bounds must reflect data ranges to prevent degenerate/faulty MFs.

7. Outlook and Research Directions

The ANFIS-PSO paradigm continues to be a focal point in interpretable AI, particularly where domain interpretability and accuracy must be balanced. Ongoing work investigates enhanced PSO variants (dynamic population, adaptive coefficient scheduling), hybridized fuzzy-deep or neuro-symbolic architectures, and domain-specific feature engineering (e.g., CSP in EEG pipelines). The transparent rule base of ANFIS, when globally tuned by PSO, offers a replicable and extensible template for nonlinear, data-driven modeling, high-fidelity control adaptation, and interpretable biomedical diagnosis. The trade-off between interpretability, robustness, and computational resource scaling remains an active area of method refinement (Aktar et al., 1 Nov 2025, Kebbati et al., 21 Sep 2025).