Evolutionary optimization framework to train multilayer perceptrons for engineering applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Conventional supervised backpropagation algorithms for training neural networks face challenges including local minimum stagnation and vanishing/exploding gradients.
• These limitations hinder convergence speed and the ability to find global minima in the loss landscape.
• Traditional methods also require extensive pre-selection of learning parameters, impacting training efficiency.

Purpose of the Study:

• To introduce and validate an evolutionary optimization framework for training multilayer perceptrons (MLPs).
• To address the limitations of conventional backpropagation algorithms in neural network training.
• To benchmark the performance of a novel dynamic group-based cooperative optimizer in training MLPs.

Main Methods:

• Developed an evolutionary optimization framework utilizing the dynamic group-based cooperative optimizer.
• Trained multilayer perceptrons using the proposed framework.
• Validated the framework on five engineering application datasets.
• Compared performance against conventional backpropagation and other evolutionary algorithms.

Main Results:

• The proposed evolutionary optimization framework demonstrated competitive performance across most examined datasets.
• The framework showed improvements in overall performance and convergence compared to traditional methods.
• For three datasets, the framework achieved performance increases of 2.7%, 4.83%, and 5.13% over the second-best optimizers.

Conclusions:

• The evolutionary optimization framework effectively trains multilayer perceptrons, offering a viable alternative to backpropagation.
• The dynamic group-based cooperative optimizer shows promise in overcoming common neural network training challenges.
• This approach enhances neural network performance and convergence, particularly for engineering applications.