paper：https://arxiv.org/abs/1912.01369
In this paper, we present NSGANetV1, a multi-objective evolutionary algorithm for NAS, extending on an earlier proof- of-principle method [20], to address the aforementioned limitations of current approaches. The key contributions followed by the extensions made in this paper are summarized below:
这篇论文提出了NSGANetV1，一种神经网络搜索的多目标遗传算法，是之前的 proof-of-principle 方法的拓展，为了解决上述现有方法的限制。这篇论文对其扩展的主要贡献总结如下：
1) NSGANetV1 populates a set of architectures to approximate the entire Pareto front in one run through customized genetic operations that recombine and modify architectural components progressively. NSGANetV1 improves computational efficiency by carefully down- scaling the architectures during the search as well as reinforcing the emerging patterns shared among past successful architectures through a Bayesian Network based distribution estimation operator. Empirically, the obtained architectures, in most cases, outperform both manually and other automatically designed architectures on various datasets.
NSGANetV1 构成了一系列结构，用来在一次自定义遗传算子运行来近似整个Pareto前沿，并逐渐修改结构组成部分。NSGANetV1 采用一个基于贝叶斯网络的分布估计算子，通过在搜索过程中仔细地向下扫描结构，加强在成功结合中显露的模式，提升了计算效率。经验来说，获得的结构，在大多场景，表现出的交互性和其他的在多样数据集上设计结构的自动性更好。
2) By obtaining a set of architectures in one run, NSGANetV1 allows designers to choose a suitable network a-posteriori as opposed to a pre-defined preference weighting of objectives prior to the search. Further post- optimal analysis of the set of non-dominated architectures often reveals valuable design principles, which is another benefit of posing NAS as a multi-objective optimization problem, as is done in NSGANetV1.
基于一次运行获得的一系列结构，NSGANetV1 允许设计者选择一个合适的网络归纳作为预定义的对目标先驱搜索的权重偏好的对立。更近的，对一系列非支配结构的后最优分析常常揭示数值射击原则，这是NAS作为多目标优化的的另一个好处，就像NSGANetV1那样。
3) From an algorithmic perspective, we extend our previous work [20] in a number of ways: (i) an expanded search space to include five more layer operations and one more option that controls the width of the network, (ii) improved encoding, mutation and crossover operators accompanying the modified search space, and (iii) a more thorough lower-level optimization process for weight learning, resulting in better and more reliable performance.
3)从算法的角度，我们扩展之前的工作，以以下几种方法（i)一个多包含五层操作和一个控制网络宽度选择 的扩展搜索空间，（ii)改进的编码，变异和交叉操作符协同修改搜索空间，（iii）一个对权重学习更详细的低层优化过程，会有一个更好更可靠的表现。
4) From an evaluation perspective, we extend our previous work [20] in two different ways: (i) adding three more tasks, including medical imaging, robustness to adversarial attacks, and car key-point estimation; and (ii) evaluating the searched architectures on five new datasets, including, ImageNet, ImageNet-V2, CIFAR- 10.1, corrupted CIFAR-10 and corrupted CIFAR-100.
4）从一个评估的角度，我们由两种方式扩展我们之前的工作:(i)增加三项任务，包括医学成像，对敌对攻击的健壮性，和一个汽车关键点估计；和（ii）在五项新数据集上估计搜索结构，包括ImageNet, ImageNet-V2, CIFAR- 10.1, corrupted CIFAR-10 and corrupted CIFAR-100
