Evolutionary learning of BMF fuzzy-neural networks using a reduced-form genetic algorithm

Area of Science:

• Computational intelligence
• Machine learning
• Artificial neural networks

Background:

• Fuzzy-neural networks (FNNs) are powerful tools for complex modeling.
• Traditional gradient-based training methods for FNNs can get stuck in local minima.
• Conventional genetic algorithms (GAs) are slow and struggle with the large number of parameters in FNNs.

Purpose of the Study:

• To propose a novel approach for optimizing FNNs using a reduced-form genetic algorithm (RGA).
• To enhance the training speed and global optimization capabilities for FNNs.
• To effectively adjust both B-spline membership function (BMF) control points and network weightings.

Main Methods:

• A reduced-form genetic algorithm (RGA) incorporating a sequential-search-based crossover point (SSCP) method was developed.
• The RGA was used to search for optimal parameters, including BMF control points and FNN weightings.
• Chromosomes were represented as real-valued vectors encoding all adjustable parameters.

Main Results:

• The proposed RGA demonstrated faster convergence in the evolutionary process compared to conventional methods.
• Simulation results showed the effectiveness of the RGA in optimizing FNNs for nonlinear function approximation.
• The RGA successfully adjusted BMF control points and FNN weightings, leading to improved performance.

Conclusions:

• The RGA offers a more efficient and effective method for training fuzzy-neural networks.
• This approach overcomes limitations of traditional gradient-based and standard GA training.
• The proposed method is well-suited for complex optimization tasks in fuzzy-neural network applications.