Training a Feedforward Neural Network Using Hybrid Gravitational Search Algorithm with Dynamic Multiswarm Particle...

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• The Gravitational Search Algorithm (GSA) is widely used for complex optimization but suffers from slow convergence and weak local search.
• Existing evolutionary learning methods often struggle with local minima trapping and poor convergence rates.

Purpose of the Study:

• To introduce and evaluate a novel hybrid algorithm, GSADMSPSO, for training Feedforward Neural Networks (FNNs).
• To address the limitations of GSA and other evolutionary methods, specifically slow convergence and local minima entrapment.

Main Methods:

• A hybrid approach combining GSA with Dynamic Multi-Swarm Particle Swarm Optimization (DMSPSO) was developed.
• The GSADMSPSO algorithm partitions the main population into smaller, stabilized subswarms with a new neighborhood plan.
• It leverages GSA's global search capability and DMSPSO's exploration-exploitation balance for enhanced performance.

Main Results:

• GSADMSPSO demonstrated superior performance compared to GSA and its variations on benchmark test functions.
• The proposed method showed significant improvements in convergence speed.
• The algorithm effectively avoided the local minima problem during FNN training.

Conclusions:

• The GSADMSPSO algorithm offers an effective solution for training FNNs, overcoming key limitations of existing methods.
• This hybrid approach enhances both the speed of convergence and the ability to escape local optima.
• GSADMSPSO presents a promising advancement in optimization techniques for machine learning applications.