Appendix for TabNAS: Rejection Sampling for Neural Architecture Search on Tabular Datasets Chengrun Y ang 1, Gabriel Bender

Due to the high costs involved, many works have proposed different methods to reduce the search cost. The first strategy is to reduce the time needed to evaluate each architecture seen during a search. The second strategy is to reduce the number of architectures we need to evaluate during a search. Resource constraints are prevalent in deep learning. Finding architectures with outstanding performance and low costs are important to both NAS research and application.