Toward learning neural network encodings for continuous optimization problems
Eric O. Scott, Kenneth Alan De Jong · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2018
To date, efforts to automatically configure problem representations for classes of optimization problems have yielded few practical results. We show that a recently proposed approach for training neural G-P maps for optimization problems yields maps that generalize poorly to translated problem instances. We propose that alternative neural architectures---especially ones that allow the number of control genes to be greater than the number of phenotypic traits---may provide a means of learning maps that are better able to generalize to new problem instances.