A Study of Generalization and Fitness Landscapes for Neuroevolution

Fitness landscapes are a useful concept for studying the dynamics of meta-heuristics. In the last two decades, they have been successfully used for estimating the optimization capabilities of different flavors of evolutionary algorithms, including genetic algorithms and genetic programming. However, so far they have not been used for studying the performance of machine learning algorithms on unseen data, and they have not been applied to studying neuroevolution landscapes. This paper fills these gaps by applying fitness landscapes to neuroevolution, and using this concept to infer useful information about the learning and generalization ability of the machine learning method. For this task, we use a grammar-based approach to generate convolutional neural networks, and we study the dynamics of three different mutations used to evolve them. To characterize fitness landscapes, we study autocorrelation, entropic measure of ruggedness, and fitness clouds. Also, we propose the use of two additional evaluation measures: density clouds and overfitting measure. The results show that these measures are appropriate for estimating both the learning and the generalization ability of the considered neuroevolution configurations.