Robust control using GA-optimized neural networks

The power of genetic algorithms is utilized in the development of robust neuro-controllers. Specifically, a genetic algorithm (GA) is used to explore the connection space of an artificial neural network (ANN) with the objective of finding a sparsely connected network that yields the best accuracy in mapping. Such sparsity is desired as it improves the generalization (robustness) capabilities of the mapping. The ANN with the GA chosen connections is then trained using a supervised mode of learning known as backpropagation of error. Two different approaches for designing robust ANN are examined. In the first approach, a GA is used to minimize the mapping error before backpropagation learning is applied. For the second approach, a GA is used to minimize the sum of second order error derivatives with respect to the ANN weights. These approaches are applied to the Space Station three-axis attitude control problem. Results observed show good robustness qualities of GA-optimized neuro-controllers