Khái niệm cốt lõi
ENAS algorithms' theoretical foundation is established by estimating the expected hitting time using a general method integrating theory and experiment.
Tóm tắt
The content introduces Evolutionary Computation-based Neural Architecture Search (ENAS) algorithms, focusing on the expected hitting time (EHT) as a key theoretical issue. The proposed method estimates EHT lower bounds for (λ+λ)-ENAS algorithms with various mutation operators. The study validates the method on NAS-Bench-101 problem, marking the first theoretical foundation for ENAS algorithms.
The structure of the content includes:
- Introduction to manually designed DNNs and motivation for NAS.
- Classification of optimization algorithms used in NAS.
- Overview of ENAS algorithms and their success in image classification.
- Theoretical background on EHT analysis in EC community.
- Various methods for analyzing EHT including fitness-level, convergence-based, drift analysis, and switch analysis.
- Challenges in applying theoretical methods to NAS due to lack of explicit fitness functions.
- Proposed CEHT-ENAS method framework integrating common configuration, search space partition, transition probability analysis, population distribution fitting, and hitting time analysis.
- Preliminaries on NAS search space, mutation-based ENAS algorithms, Markov chain modeling.
- Detailed steps for fitting population distribution based on experimental data collection and surface fitting techniques.
Thống kê
この論文は、(λ+λ)-ENASアルゴリズムにおけるEHTの下限値を推定するための一般的な方法を提案しています。
この研究では、NAS-Bench-101問題で提案された手法の妥当性を示しています。