A memetic dynamic coral reef optimisation algorithm for simultaneous training, design, and optimisation of artificial...

Area of Science:

• Computational Intelligence
• Machine Learning
• Artificial Intelligence

Background:

• Gradient descent algorithms like Backpropagation (BP) are standard for Artificial Neural Network (ANN) optimization but can get stuck in local optima.
• Designing ANN architecture requires careful consideration, and traditional methods often struggle with complex optimization landscapes.

Purpose of the Study:

• To propose a novel memetic training method for simultaneously learning ANN structure and weights.
• To introduce and evaluate three versions of a Coral Reef Optimization (CRO) based algorithm for ANN training.
• To demonstrate the superiority of the proposed M-DSCRO algorithm over existing methods.

Main Methods:

• Development of three memetic training algorithms based on Coral Reef Optimization (CRO): Memetic CRO, Memetic SCRO (M-SCRO), and Memetic Dynamic SCRO (M-DSCRO).
• Adaptation of CRO algorithms with specific operators tailored for ANN design.
• Evaluation of algorithm performance on 40 diverse classification datasets.

Main Results:

• The M-DSCRO algorithm demonstrated superior performance compared to the other two proposed CRO versions across most datasets.
• M-DSCRO achieved better overall accuracy and improved performance on minority classes when compared to four state-of-the-art methods.
• The dynamic statistical approach in M-DSCRO enhanced the evolutionary process for ANN optimization.

Conclusions:

• The proposed M-DSCRO algorithm offers a robust and effective approach for optimizing Artificial Neural Network structure and weights.
• Memetic Coral Reef Optimization presents a viable alternative to traditional gradient-based methods for complex ANN training tasks.
• M-DSCRO shows significant potential for improving classification accuracy, especially for imbalanced datasets.