AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

Optimizing Network Simulation: Enhancing Performance Prediction Accuracy via Neural Architecture Search

ShaoChen He, Zirui Zhuang, Haifeng Sun, Xiaoyuan Fu, Qi Qi, Lei Zhang, Jianxin Liao, Jingyu Wang

ICML 2026 · Architectures and efficiency › Neural architecture search · approach: Latency-aware architecture search

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Wasserstein loss to MSE training (ablation)ANAS · DeepQueueNet switch dataset (test)p99Wn (x1e-3): 18.1 -> 17.0Improved
add Wasserstein loss to MSE training (ablation)ANAS · DeepQueueNet switch dataset (test)MSE (x1e-5): 5.32 -> 5.14Improved
add similarity-guided evaluation and dynamic warm-up (ablation)ANAS · DeepQueueNet switch dataset (test)MSE (x1e-5): 5.78 -> 5.14Improved
use hybrid attention-recurrent encoder-decoder search space instead of cell-based space (ablation)ANAS · DeepQueueNet switch dataset (test)MSE (x1e-5): 5.94 -> 5.14Improved
ANAS search strategy vs double-budget random search in same space and loss (ablation)ANAS · DeepQueueNet switch dataset (test)MSE (x1e-5): 5.64 -> 5.14 (32h -> 16h)Improved
replace manual DeepQueueNet with ANAS-discovered architectureDeepQueueNet · DeepQueueNet switch dataset (test)MSE (x1e-5): 6.93 -> 5.14Improved
use general-purpose Transformer instead of ANAS architectureTransformer · DeepQueueNet switch dataset (test)p99Wn (x1e-3): 17.0 -> 22.1Hurt
use ENAS (standard weight-sharing NAS) instead of ANAS under distribution shiftENAS · Roblox trace end-to-endp99Delay Wn: 0.087 -> 0.857Hurt
lower inter-architecture similarity between training and evaluated architecture (ablation)weight-sharing supernet evaluation · Supernet evaluation L1 lossmean L1 loss (x1e-3): 9.6 -> 13.7Hurt
use ANAS in High Load scenario vs DeepQueueNetANAS · High Load (HL) end-to-endavgDelay Wn: 0.034 -> 0.026Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy