Optimizing Extreme Learning Machine Neural Networks Using Differential Evolution

Differential Evolution (DE), as an efficient swarm intelligence optimization technique, combined with Extreme Learning Machine (ELM) neural networks, provides a novel approach to address traditional neural network training challenges. The core characteristic of ELM lies in randomly generating weights and biases from the input layer to the hidden layer, requiring only analytical computation for output layer weights. However, this randomness can lead to unstable network performance, which is precisely where the value of introducing Differential Evolution becomes apparent.

Differential Evolution iteratively optimizes populations through three core operations: mutation, crossover, and selection. When applied to ELM, it primarily optimizes three key aspects: first, the initial weight matrix from input layer to hidden layer; second, the bias parameters of hidden layer neurons; and finally, the optimal activation function type. This optimization process significantly enhances the network's generalization capability and avoids performance fluctuations caused by random initialization.

In MATLAB implementation, special attention should be paid to algorithm parameter settings: population size is typically set to 5-10 times the number of parameters to be optimized, mutation factor F should take values between 0.5-1.0, and crossover probability CR is recommended to be adjusted within the range of 0.3-0.9. The design of the fitness function is crucial for the entire optimization process, generally using mean squared error or classification accuracy on the validation set as evaluation metrics. Key implementation considerations include using vectorized operations for efficient matrix computations and properly handling boundary conditions during mutation operations.

The advantage of this hybrid method lies in: DE's global search capability compensates for the limitations of ELM's random initialization, while ELM's high-speed training characteristics ensure optimization process efficiency. In practical applications, this method is particularly suitable for regression and classification problems on small to medium-sized datasets, maintaining ELM's training speed advantage while significantly improving model prediction accuracy and stability. The implementation typically involves creating separate functions for DE operations (mutation, crossover, selection) and integrating them with ELM's weight calculation procedures.