An analysis of genetic algorithm training of a wind tunnel neural network control system

Wind tunnel testing requires an accurate and speedy transition between data points. Efficient transition decreases testing time and project cost. Addition of neural network prediction to standard feedback controllers reduces tunnel condition settling times. Moreover, change in tunnel operating characteristics due to change in test article configuration is compensated by neural prediction that is refined by re-training. This work compares back-propagation and genetic algorithm training methods for achieving trained neural predictors. A computer wind tunnel model provides a common data set for training method analysis. Results show that hybrid genetic algorithm training and back-propagation training produce similar Mean Squared Error (MSE). However, genetic algorithm training requires more time for convergence. Genetic operations provide the advantage of producing a global solution for the neural network. For relatively linear wind tunnel model training data, back-propagation is preferred.