Evolutionary artificial neural networks by multi-dimensional particle swarm optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Designing Artificial Neural Networks (ANNs) often requires predefining network architecture, a limitation in complex search spaces.
• Existing swarm optimization techniques struggle when the optimal number of dimensions (network complexity) is unknown.
• Automatic design of ANNs is crucial for optimizing performance and generalization across diverse problems.

Purpose of the Study:

• To propose a novel multi-dimensional Particle Swarm Optimization (MD PSO) technique for the automatic design of ANNs.
• To enable ANNs to evolve towards optimal configurations within an architecture space without a priori dimension specification.
• To remove the drawback of fixed dimension settings inherent in traditional swarm optimizers.

Main Methods:

• The study employs a multi-dimensional Particle Swarm Optimization (MD PSO) technique.
• Particles are restructured to perform inter-dimensional passes, enabling simultaneous search for positional and dimensional optima.
• Network configurations and parameters are encoded into particles for optimization in both error and architecture spaces.

Main Results:

• MD PSO successfully evolves optimal or near-optimal Artificial Neural Networks.
• The technique demonstrates superior generalization capabilities on benchmark problems.
• MD PSO naturally favors simpler, compact network configurations when performance is comparable to complex ones.

Conclusions:

• The proposed MD PSO technique effectively automates ANN design by optimizing both network parameters and architecture dimensions.
• This method overcomes the limitations of fixed-dimension optimization in swarm intelligence.
• The approach yields efficient and well-generalizing ANN models, offering a ranked list of configurations for practical application.