Prediction of neural network performance by phenotypic modeling

摘要: Surrogate models are used to reduce the burden of expensive-to-evaluate objective functions in optimization. By creating which map genomes values, these can estimate performance unknown inputs, and so be place expensive functions. Evolutionary techniques such as genetic programming or neuroevolution commonly alter structure genome itself. A lack consistency genotype is a fatal blow data-driven modeling techniques: interpolation between points impossible without common input space. However, while dimensionality genotypes may differ across individuals, many domains, controllers classifiers, output remains constant. In this work we leverage insight embed differing neural networks into same To judge difference behavior two networks, give them both sequence, examine output. This difference, phenotypic distance, then situate space, allowing us produce surrogate predict regardless topology. robotic navigation task, show that trained using embedding perform well better those on weight values fixed topology network. We establish promising flexible approach enables even for representations undergo structural changes.