Mind Evolutionary Algorithm Optimized BP Neural Network

Application Background Sun Chengyi and colleagues proposed the Mind Evolutionary Algorithm (MEA) in 1998, which is recognized as an effective optimization method. This chapter provides a detailed explanation of MEA's fundamental principles and demonstrates its implementation in MATLAB environment through a nonlinear function fitting case study. Key Technologies 1. Training/Test Set Generation: The generation of training and testing datasets is crucial for algorithm evaluation. In MATLAB implementation, this typically involves creating input-output pairs using functions like rand() for random sampling or linspace() for evenly distributed points, followed by data normalization using mapminmax() function. 2. Initial Population Generation: Proper initialization significantly impacts algorithm performance. The implementation involves defining population size and chromosome length, then using unifrnd() function to generate individuals within specified bounds while ensuring diversity through random initialization techniques. 3. Subpopulation Convergence Operation: This crossover operation enhances population diversity and search capability. The MATLAB code typically implements tournament selection followed by simulated binary crossover (SBX) with controlled distribution index, maintaining population size while exploring new solution spaces. 4. Subpopulation Dissimilation Operation: This mutation operation introduces new individuals and explores uncharted search regions. The implementation uses polynomial mutation with adaptive mutation rates, where mutation probability decreases as generations progress to balance exploration and exploitation. 5. Optimal Individual Analysis: Each generation requires fitness evaluation and elite selection. The code implements sorting algorithms like sort() function to identify best-performing individuals, with elite preservation ensuring best solutions carry forward to next generations. 6. BP Neural Network Training: For the nonlinear function fitting case, MEA optimizes BP neural network parameters. The implementation involves configuring network architecture using newff() function, where MEA optimizes connection weights and biases through fitness function minimization, typically using mean squared error (MSE) as objective function. 7. Simulation Testing and Result Analysis: Final validation includes comprehensive testing with performance metrics calculation. The MATLAB implementation compares predicted outputs with actual values, generates convergence curves using plot() function, and performs statistical analysis using functions like mean() and std() to evaluate algorithm robustness and efficiency.